A method for constructing a communication network video ringback tone XR application

By using Unity environment rendering components, distributed resource management and network monitoring technology in the video ringtone system, the problems of real-time rendering and network environment impact are solved, and efficient and personalized video ringtone experience and system stability are achieved.

CN118118595BActive Publication Date: 2024-11-15CHINA UNICOM WO MUSIC & CULTURE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410383248.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-29
Publication Date
2024-11-15
Estimated Expiration
2044-03-29

AI Technical Summary

Technical Problem

Existing video ringtone systems face high computing power and response speed requirements, increased system complexity, stability and scalability challenges, and rendering quality degradation caused by uncertainty in network environments.

Method used

Adopt the Render Streaming component based on the Unity environment for digital human and scene rendering, deploy distributed resource management module to monitor the computing node load in real time, adjust the rendering task priority according to performance indicators and user experience, and pre-fetch data through Nagios monitoring network bandwidth and latency and synchronize, use the Coturn server to establish audio and video call links, and combine the DTLS encryption protocol and WebRTC protocol to optimize rendering stream quality.

Benefits of technology

It realizes efficient real-time rendering, provides a highly personalized video ringtone experience, ensures service continuity and user experience, improves rendering task processing efficiency and quality, and improves system robustness and resource utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118118595B_ABST
    Figure CN118118595B_ABST
Patent Text Reader

Abstract

The present application provides a method for constructing a communication network video ringback tone XR application, including: based on the Render Streaming component of the unity environment, rendering and processing the digital human and scene in the video ringback tone on the rendering server; deploying a distributed resource management module, monitoring the load data of each computing node of the rendering server in real time, judging whether it is a faulty node, and reallocating the rendering tasks corresponding to the faulty nodes to non-faulty nodes, the rendering tasks including digital human rendering and scene rendering; establishing an audio and video call link between the rendering server and the user terminal through the Coturn server, and transmitting the rendered digital human and scene by the audio and video call link after the rendering task is completed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for constructing a communication network video ringback tone XR application. Background Art

[0002] In the current communication and media fields, video ringback tone has become a popular personalized service that allows users to display video content during call waiting to provide a richer interactive experience. With the rapid development of virtual reality, augmented reality, and extended reality technologies, expectations for video ringback tone services have expanded from simple video playback to real-time interaction and highly personalized digital human and scene rendering. This new type of video ringback tone XR application poses a series of technical challenges, especially in terms of real-time, personalization, stability, and scalability. First, the real-time rendering problem is one of the core challenges in building high-quality video ringback tone XR applications. Traditional video ringback tone usually only contains pre-recorded video content without the need for real-time rendering. However, the real-time rendering of digital humans and scenes requires the system to be able to respond to user interactions instantly and dynamically adjust the rendering content according to the user's personalized needs, which places extremely high demands on the computing power and response speed of the rendering system. Secondly, highly personalized video ringback tone services increase the complexity of the system. Users want to show their personality and preferences through video ringback tone, which requires the system to be able to intelligently adjust the rendering strategy and parameters according to the specific needs of users, such as personal preferences, device capabilities, network conditions, and other factors. This not only involves complex user portrait analysis, but also requires the system to have flexible rendering parameter adjustment capabilities. Thirdly, the stability and scalability of the system are also key challenges. With the increase in the number of users and the complexity of rendering tasks, maintaining high availability and high performance of the system has become a major challenge. How to effectively manage computing resources, how to handle fault recovery of computing nodes, and how to dynamically adjust resource allocation based on real-time monitoring data are all technical problems that need to be solved. Finally, the uncertainty of the network environment has a direct impact on the user experience of video ringtone XR applications. Since XR content usually requires higher data transmission rates and lower latency, network bandwidth limitations and unstable network connections may cause delays or degradation in the quality of rendered content. Therefore, how to ensure efficient data transmission and rendering quality in an unstable network environment is another technical problem that needs to be solved. Summary of the invention

[0003] The present invention provides a method for constructing a communication network video ringback tone XR application, which mainly includes:

[0004] Based on the Render Streaming component of the Unity environment, the digital human and scene in the video ringback tone are rendered on the rendering server;

[0005] Deploy a distributed resource management module to monitor the load data of each computing node of the rendering server in real time, determine whether it is a faulty node, and reallocate the rendering tasks corresponding to the faulty node to the non-faulty node. The rendering tasks include digital human rendering and scene rendering.

[0006] Based on the real-time monitoring data of each computing node of the rendering server, the performance of each computing node of the rendering server is evaluated, and the performance index of each computing node of the rendering server is determined by combining the multi-dimensional performance indicators of each computing node, including processing speed, stability and error rate;

[0007] According to the deadline of the task in the video ringtone rendering task, the historical digital human rendering data and the scene rendering data, and the urgency, complexity and estimated rendering time of each rendering task are determined, a task priority model is established, and the rendering tasks are divided into critical rendering tasks and non-critical rendering tasks. According to the performance index of each computing node of the rendering server and the priority of each rendering task, a node allocation score evaluation formula is constructed to generate a rendering task allocation strategy;

[0008] Adjust the priority of non-critical rendering tasks based on user experience indicators, including response time and playback smoothness. According to the user attributes of the callee, determine the preference of digital human rendering and scene rendering in the callee's video ringback tone, and adjust the digital human rendering parameters and scene rendering parameters based on the preference of digital human rendering and scene rendering in the callee's video ringback tone.

[0009] Monitor network bandwidth utilization and latency through Nagios, perform data pre-fetching and data synchronization based on monitoring results, use LRU and FIFO to manage cache servers corresponding to each computing node, dynamically allocate rendering tasks based on the status of the cache server, and continuously optimize the network environment and cache server configuration based on cache server load and performance indicators;

[0010] Regularly check the health status of each computing node of the rendering server. If the response time of a computing node exceeds a preset threshold, it is identified as an unhealthy node and triggers task reallocation, including making decisions based on performance evaluation data and task classification data to migrate tasks of unhealthy nodes to other healthy nodes with matching performance.

[0011] An audio and video call link is established between the rendering server and the user terminal through the Coturn server, and the rendered digital human and scene are transmitted by the audio and video call link after the rendering task is completed;

[0012] Based on the historical performance data of each computing node of the rendering server, a node performance prediction model is constructed to obtain the load and performance changes of each computing node of the rendering server in a preset time period in the future, determine the performance bottleneck of each computing node, and redistribute the rendering tasks in advance.

[0013] In one implementation, the Render Streaming component based on the unity environment performs rendering processing on the digital human and scene in the video ringback tone on the rendering server, including:

[0014] By comparing the encoding efficiency, visual quality and bandwidth consumption, the performance of different encoders under the same network conditions is evaluated, the most suitable encoder is selected, and the encoder configuration is updated regularly. Different encoders include H.264 or H.265. The network quality monitoring tool Wireshark is used to obtain real-time data on packet loss rate, delay and bandwidth. If the real-time data shows that the network condition causes the quality of RenderStreaming to decline, the transmission strategy is adjusted, including changing the encoding quality or enabling a resolution lower than the preset resolution threshold. Based on the complexity of the digital human model and scene in the video ringback tone, the current GPU performance and CPU performance of the rendering server are evaluated to determine whether the preset minimum standards for real-time rendering of digital human models and scenes are met. If the server performance does not meet the minimum standards for real-time rendering of digital human models and scenes, the hardware is upgraded to improve the rendering capability of the server. According to the client browser's support for WebRTC, including the browser version, platform characteristics and support for the core functions of WebRTC, the best WebRTC compatibility solution is selected, and the rendering effect and fluency on different browsers are tested. The core functions of WebRTC include audio and video calls and data channels. Design a WebSocket-based t's real-time communication protocol is used to encode user input and transmit it to the rendering server in real time. The rendering server adjusts the state or behavior of the digital human model according to the received instructions; UnityProfiler is used to analyze the performance bottleneck of the scene and implement scene optimization measures. The scene optimization measures include eliminating objects that are not in the field of view and using Shader to reduce the number of dynamic light sources. The scene performance bottleneck includes objects that are out of sight and dynamic light sources that exceed the preset number in the scene; according to the position and role of the digital human in the scene, the Camera parameters in Unity, including the viewing angle and field of view, are configured and adjusted to obtain the best rendering viewing angle and visual effect; if the rendering content contains pre-recorded video, the VideoPlayer component in Unity is used to control the video playback, and the pre-recorded video is synchronously sent to the client as part of the rendering stream; the video stream data is encrypted through the DTLS encryption protocol, and the encryption mechanism is tested to determine whether the encryption affects the video stream transmission and rendering efficiency, and the encryption algorithm and configuration are adjusted accordingly; the real-time monitoring tool Prometheus is integrated into Unity to monitor the performance of the rendering stream in real time, including delay and frame loss, and the encoding settings or network configuration are adjusted according to the monitoring results to optimize the rendering stream quality.

[0015] In one embodiment, the distributed resource management module is deployed to monitor the load data of each computing node of the rendering server in real time, determine whether it is a faulty node, and reallocate the rendering tasks corresponding to the faulty node to non-faulty nodes, wherein the rendering tasks include digital human rendering and scene rendering, including:

[0016] By deploying the distributed resource management module Collectd, the load data of each computing node is obtained on each computing node of the rendering server. The load data includes CPU utilization, memory usage, and network bandwidth. The obtained load data is sent to the central monitoring server, and the message middleware Kafka is used for real-time data transmission. The monitoring tool Prometheus is used to process and store the obtained load data to obtain the real-time load of each computing node. The data visualization tool Grafana is used to convert the processed performance data into charts and dashboards to monitor the status of the computing nodes of the rendering server in real time. If the monitoring data shows that the load data of any node exceeds the predetermined load, the load data is automatically forwarded to the central monitoring server. If the load threshold is exceeded, the alarm and notification mechanism is triggered, and the node is judged to be a faulty node; the resource management and scheduling framework Kubernetes is used to dynamically allocate computing node resources based on real-time monitoring data; cluster status management information is obtained to monitor the health status of each node in the cluster, and the automation tool Ansible is used to perform configuration management and maintenance tasks; if a computing node failure is detected, the fault tolerance mechanism is activated to reallocate the rendering tasks on the node to non-faulty nodes. The rendering tasks include digital human rendering and scene rendering; the log files generated by the rendering server are obtained through Logstash, and historical performance data and event logs are regularly analyzed to optimize resource scheduling strategies and improve fault recovery processes.

[0017] In one embodiment, the real-time monitoring data of each computing node of the rendering server is used to evaluate the performance of each computing node of the rendering server, and the performance index of each computing node of the rendering server is determined by combining the multi-dimensional performance indicators of each computing node, wherein the multi-dimensional performance indicators include processing speed, stability and error rate, including:

[0018] The key performance indicators of each computing node of the rendering server are regularly obtained through the real-time monitoring tool Prometheus, including data on CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput; if the CPU or GPU usage in the obtained data continues to reach the preset peak value, it is determined whether there is uneven task distribution or hardware performance limitations; the task processing speed of each node of the rendering server is evaluated through statistical data on task completion time. For nodes with processing time longer than the preset time threshold, the hardware and server configuration are checked to determine the cause of the delay; through the load data of each computing node of the rendering server, it is determined that each computing node of the rendering server is running at a load higher than the preset load threshold Stability: If the load of the computing node is continuously higher than the preset load threshold, the node is judged to be at risk of overload; the log files generated by the rendering server are obtained through Logstash, and the data on the number of hardware failures, server crashes, and abnormal program exits are obtained from the log files to evaluate the stability of each node; the number of rendering errors and log error information are obtained from the log files generated by the rendering server, and the error rate of the node is calculated based on the number of rendering errors and log error information. If the error rate of the node increases, it is inferred that there is a problem with the hardware or software configuration; according to the multi-dimensional performance indicators, including processing speed, stability, and error rate, weights are assigned to each multi-dimensional performance indicator, and the performance index calculation formula is adopted, PI = (ω v ×S v )+(ω s ×S s )+(ω e ×S e ), calculate the performance index of each computing node, where PI represents the performance index, ω v represents the weight of processing speed, S v represents the score of processing speed, ω s represents the weight of stability, S s represents the stability score, ω e represents the weight of the error rate, S e The score represents the error rate. According to the calculated performance index, the computing nodes are sorted, and the nodes with performance lower than the preset performance threshold are identified, and the resource scheduling measures are adjusted.

[0019] In one embodiment, the method determines the urgency, complexity and estimated rendering time of each rendering task according to the deadline of the task in the video ringback tone rendering task, the historical digital human rendering data and the scene rendering data, establishes a task priority model, divides the rendering tasks into key rendering tasks and non-key rendering tasks, and constructs a node allocation score evaluation formula according to the performance index of each computing node of the rendering server and the priority of each rendering task, and generates a rendering task allocation strategy, including:

[0020] Get the rendering task schedule, determine the deadline for the video ringtone rendering task, and if the difference between the deadline and the current time is lower than the preset time threshold, mark the urgency of the rendering task as high; adjust the urgency rating according to the customer's requirements for task completion time, and increase the urgency of rendering tasks with time requirements higher than the preset time threshold; use database query to obtain historical digital human rendering data and scene rendering data. Digital human rendering data includes digital human model detail level, digital human texture resolution, digital human material complexity, and digital human dynamic effect requirements. Scene rendering data includes scene size and scene environment effects. Use decision tree algorithm to train the model and evaluate the complexity of the current rendering task. The complexity includes high, medium and low. Get hardware capability parameters, combine with rendering setting information, and use Blender simulation to estimate the expected rendering time of each rendering task. Hardware capability parameters include CPU and GPU performance indicators. Rendering setting information includes resolution and sampling rate; Get the availability of current hardware resources and human resources. If hardware resources and human resources are tight, for tasks with complexity higher than the preset complexity threshold and estimated rendering time longer than the preset rendering time threshold, task, improve the priority of the rendering task; obtain the dependency relationship between the rendering tasks, if the current rendering task is a prerequisite for multiple subsequent rendering tasks, then improve the priority of the rendering task; according to the urgency, complexity, estimated rendering time and dependency of the task, use the recurrent neural network algorithm to train the model, establish a task priority model, and comprehensively score each rendering task; according to the comprehensive scoring result, sort all rendering tasks, and divide them into key rendering tasks and non-key rendering tasks according to the score, where the comprehensive score higher than the preset scoring threshold is divided into key rendering tasks, and the comprehensive score lower than the preset scoring threshold is divided into non-key rendering tasks; obtain the performance index, load condition and rendering task priority of each computing node of the rendering server, construct a node allocation score evaluation formula, node allocation score = λ×performance index-μ×load condition, allocate the best computing node to each rendering task, and generate a rendering task allocation strategy, where λ and μ are weight coefficients used to balance the impact of performance and load; according to the generated rendering task allocation strategy, allocate the rendering task to the selected computing node, and give priority to the key rendering task.

[0021] In one embodiment, the priority of non-critical rendering tasks is adjusted based on user experience indicators, where the user experience indicators include response time and playback smoothness, and the preference of digital human rendering and scene rendering in the video ringback tone of the call object is determined according to the user attribute of the call object, and digital human rendering parameters and scene rendering parameters are adjusted based on the preference of digital human rendering and scene rendering in the video ringback tone of the call object, including:

[0022] Monitor user experience indicator data in real time, including response time and playback fluency; use support vector machine algorithm to train models based on the acquired user experience indicator data to determine whether there is a downward trend in user experience quality; if there is a downward trend in user experience quality, implement a dynamic priority adjustment mechanism to dynamically reduce the priority of non-critical rendering tasks being processed based on real-time user experience indicators, and give priority to rendering tasks that affect user experience; use decision tree algorithm to train models based on historical user experience data, predict changes in user experience indicators, and adjust task priorities in advance before user experience indicators decrease; during the execution of rendering tasks, continuously monitor the execution status and user experience indicators of each rendering task, and adjust the rendering task allocation strategy based on real-time data; obtain the device performance parameters of the call object, including the number of CPU cores, memory size, and GPU type, through the device performance detection library navigator.hardwareConcurrency; set a performance baseline based on device performance parameters to determine whether the device can carry high, medium, and low levels of rendering complexity; use the network monitoring tool SpeedtestAPI to obtain the network of both parties on the call Bandwidth and connection speed data. If the network bandwidth of either party of the call is lower than a predetermined threshold, the resolution and frame rate of the video call are reduced. User preference options are added to the application settings, including video resolution preferences and quality priority or speed priority options, and the user preference settings are obtained using user profiles stored locally or in the cloud. Based on user preferences and environmental sensor data, the brightness and background noise level of the environment in which the callee is located are determined, and the brightness and volume settings of the video ringback tone are adjusted. Based on the user attributes and historical usage data of the callee, a decision tree algorithm is used to train the model to predict the video ringback tone style preference of the callee. User attributes include age, gender, and hobbies. Based on the video ringback tone style preference of the callee, the rendering style of the digital human and the scene is adjusted. Based on the device performance, network conditions, video ringback tone style preference, and environmental data of the callee, a recurrent neural network algorithm is used to train the model to obtain the optimal digital human rendering parameter settings and scene rendering parameter settings for the call. The digital human rendering parameters include the digital human model detail level, digital human texture resolution, digital human material complexity, and digital human dynamic effect requirements. The scene rendering parameters include scene size and scene environmental effects.

[0023] In one embodiment, the network bandwidth utilization and latency are monitored by Nagios, data pre-fetching and data synchronization are performed based on the monitoring results, LRU and FIFO are used to manage the cache servers corresponding to each computing node, and rendering tasks are dynamically allocated based on the status of the cache servers, and the network environment and cache server configuration are continuously optimized based on the cache server load and performance indicators, including:

[0024] Through the network monitoring tool Nagios, the bandwidth utilization and latency of the current network are continuously detected. If changes in the network environment are detected, the pre-fetch mechanism is triggered. Changes in the network environment include bandwidth utilization increases greater than the preset increase threshold or latency increases greater than the preset increase threshold. According to the pre-fetch trigger conditions, a data pre-fetch strategy is adopted to obtain historical access patterns and user behavior data related to the task, and a decision tree algorithm is used for model training to determine the resource data that should be pre-loaded into the cache. If it is determined that resource data needs to be pre-fetched, a timestamp mechanism is applied to synchronize the source data with the data in the cache. After completing data pre-fetching, the data expiration strategy in the cache strategy is used to manage the data in the cache. The data expiration strategy includes least used LRU and first-in-first-out FIFO. Based on the node allocation score evaluation formula, the rendering task is dynamically allocated according to the resource status and task requirements of the cache server. If the specified server resources are detected, If the load is higher than the preset second load threshold, the rendering task to be assigned will be reallocated to the server whose load is lower than the preset third load threshold; the performance of the cache server is evaluated according to the hit rate and service quality indicators in the performance indicators of the cache server. If the performance of the cache server does not meet the standards, the cache server performance level is expanded by adjusting the cache size or increasing the number of servers; based on the proportion of time when the cache server operates normally, the cache server availability indicator is evaluated, and the response time data is obtained, where the response time is the time from the user initiating the request to the cache server responding; the user experience quality of the user during the execution of the rendering task is continuously evaluated through the response time and cache server availability indicators. If a user experience problem is detected, the network environment and cache server configuration are optimized; a fault recovery strategy is applied. When a cache server fails, switch to a backup server or reallocate resources. The fault recovery strategy includes a backup mechanism and a fault tolerance mechanism.

[0025] In one embodiment, the health status of each computing node of the rendering server is regularly detected. If the response time of the computing node exceeds a preset threshold, it is identified as an unhealthy node and task reallocation is triggered, including making a decision based on performance evaluation data and task classification data, and migrating the tasks of the unhealthy node to other healthy nodes with matching performance, including:

[0026] Implement a scheduled task mechanism, set rendering task detection intervals, detect the health detection indicators of each computing node of the rendering server, and evaluate the health status of each computing node. The health detection indicators include CPU usage and memory usage; regularly obtain the response time data of each node, and use the monitoring tool Prometheus combined with Grafana to determine whether the node is within the preset response time range; for nodes with abnormal response time or other abnormal health detection indicators, perform retry detection. If the number of consecutive detection failures is greater than the preset threshold, the node is determined to be an unhealthy node; obtain the error log of the unhealthy node, use the abnormal information to assist in determining the fault type of the node, and determine the corresponding fault repair measures; obtain the task related to the unhealthy node from the task queue and the mapping information between the task and the computing node service information and evaluate the current status of the task; for each task to be migrated, based on the resource requirements of the migration task and the resource supply capacity of the node, calculate the performance matching degree between the migration task and each normally operating node, and the resource requirements include CPU intensive and memory intensive; based on the urgency of the task, resource requirements and user needs, use the decision tree algorithm to train the model and evaluate the priority of all tasks to be migrated, including high priority, medium priority and low priority; define the migration strategy according to the priority of the task and the performance matching metric, and the migration strategy includes determining the order of migration of the migration task and selecting the best target node. The migration task with a high priority is migrated first and migrated to the node with the highest performance matching degree; according to the migration strategy, migrate the tasks of the unhealthy node to other healthy nodes with matching performance.

[0027] In one embodiment, an audio and video call link is established between the rendering server and the user terminal through the Coturn server, and the audio and video call link transmits the rendered digital human and scene after the rendering task is completed, including:

[0028] According to the network environment of the rendering server and the user terminal, the STUN service is configured through the Coturn server to obtain the public IP and port information of the rendering server and the user terminal, and determine whether there are NAT or firewall restrictions; if the user terminal and the rendering server are both behind NAT, the TURN service of the Coturn server is used to establish a relay to transmit data to devices behind different NATs; the security protocol DTLS is used to encrypt the audio and video data stream to obtain an audio and video transmission solution that protects the communication link; the WebRTC protocol is used to establish a real-time audio and video call link between the rendering server and the user terminal to evaluate the bandwidth and delay of the link, and optimize the quality of audio and video transmission according to the bandwidth and delay of the link; according to the transmission quality of the real-time call link, the audio and video attributes are adjusted, including reducing the resolution or frame rate, to obtain an appropriate The best call effect for the current network environment; judge the processing power of the rendering server through the performance parameters of the rendering server, and determine the complexity of the rendering task based on the processing power of the rendering server. The performance parameters of the rendering server include CPU performance and GPU performance; obtain the device type and operating system information of the user terminal, evaluate the compatibility and display effect of the rendered content on the user terminal, and determine the best rendering parameters for the user device; establish a session management mechanism for audio and video calls through the WebRTC session control protocol SessionDescriptionProtocol, and synchronously transmit the rendering tasks of the digital human model and the scene on the call link; after the rendering task is completed, transmit the rendered digital human and scene to the user terminal through the established secure and encrypted audio and video call link to achieve real-time audio and video interaction.

[0029] In one embodiment, the node performance prediction model is constructed based on the historical performance data of each computing node of the rendering server to obtain the load and performance changes of each computing node of the rendering server in a future preset time period, determine the performance bottleneck of each computing node and perform the redistribution of rendering tasks in advance, including:

[0030] Obtain key performance indicator data of each computing node of the rendering server, including data on CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput; identify the demand for computing resources and node resource consumption patterns of different types of rendering tasks based on historical rendering task data. Different types of rendering tasks include CPU-intensive, GPU-intensive, and memory-intensive tasks; organize the obtained key performance indicator data in timestamp sequence, use the ARIMA algorithm for model training, predict the key performance indicator values ​​in the future preset time period, and determine the periodic fluctuations and time dependence of node performance based on the predicted trend line, seasonal component, and residual; determine the correlation between task execution mode and node performance based on historical task execution data using the method of calculating the Pearson correlation coefficient. Historical task execution data includes task duration, resource consumption, and queue waiting time; obtain system event data from system maintenance logs, hardware replacement records, and abnormal event logs. For each obtained system event, associate the computing node performance indicator data in the corresponding time period. According to the changes in performance indicators before and after the system event, identify the impact of the system event on the computing node The system event data includes the date and time of maintenance, description of maintenance activities, replacement parts and reasons for hardware replacement, time, type and impact range of abnormal events; obtain key hardware configuration information of each computing node of the rendering server, including the number of CPU cores, CPU model, GPU model and computing power, memory capacity, disk type, disk capacity and network interface type; associate the obtained hardware configuration information with the performance indicator data of the node, use the variance analysis method to determine the correlation between the hardware configuration parameters and the node performance indicators, and identify the hardware configuration factors whose impact on the node performance is greater than the preset threshold; according to the hardware configuration, performance indicators, task characteristics and system event log data, use the random forest algorithm to train the model, build a node performance prediction model, and predict the performance changes of the computing node in the future preset time period; according to the prediction results of the node performance prediction model, identify potential performance bottleneck areas, where the performance bottleneck is a node with degraded performance; based on the prediction results of the node performance prediction model and the performance bottleneck prediction, adjust the distribution of rendering tasks in advance. If a node is predicted to face a performance bottleneck, reallocate the rendering tasks in advance.

[0031] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0032] The present invention provides a method for constructing a communication network video ringback tone XR application. Through advanced distributed management, efficient real-time rendering of digital characters and scenes in video ringback tones is achieved, ensuring that rendering tasks can be completed quickly and accurately. In addition, the system can dynamically adjust the rendering content according to the user's preferences and needs to provide a highly personalized video ringback tone experience. In the event of a computing node failure, tasks can be quickly reallocated to ensure that the continuity of the service is not affected. By monitoring the network status and user device performance, the rendering parameters are intelligently adjusted to adapt to various network environments and device limitations, thereby ensuring a high-quality user experience. Overall, the present invention significantly improves the processing efficiency and quality of rendering tasks through intelligent resource management and optimization, bringing users a smooth, high-definition video ringback tone viewing experience, while improving the robustness and resource utilization efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a flowchart of a method for constructing a communication network video ringback tone XR application.

[0034] Figure 2 A schematic diagram of a method for constructing a communication network video ringback tone XR application of the present invention.

[0035] Figure 3 It is another schematic diagram of a method for constructing a communication network video ringback tone XR application of the present invention. DETAILED DESCRIPTION

[0036] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0037] like Figure 1-3 In this embodiment, a method for constructing a communication network video ringback tone XR application may specifically include:

[0038] S101. Based on the Render Streaming component of the unity environment, the digital human and the scene in the video ringback tone are rendered on the rendering server.

[0039] By comparing encoding efficiency, visual quality and bandwidth consumption, evaluate the performance of different encoders under the same network conditions, select the most suitable encoder, and regularly update the encoder configuration. Different encoders include H.264 or H.265. Use the network quality monitoring tool Wireshark to obtain real-time data on packet loss rate, latency and bandwidth. If the real-time data shows that the network condition causes the quality of RenderStreaming to decline, adjust the transmission strategy, including changing the encoding quality or enabling a resolution lower than the preset resolution threshold. Based on the complexity of the digital human model and scene in the video ringback tone, evaluate the current GPU performance and CPU performance of the rendering server to determine whether it meets the preset minimum standards for real-time rendering of digital human models and scenes. If the server performance does not meet the minimum standards for real-time rendering of digital human models and scenes, upgrade the hardware to improve the rendering capability of the server. According to the client browser's support for WebRTC, including browser version, platform characteristics and support for WebRTC core functions, select the best WebRTC compatibility solution and test the rendering effect and fluency on different browsers. WebRTC core functions include audio and video calls and data channels. Design a real-time communication protocol based on WebSocket to encode user input and transmit it to the rendering server in real time. The rendering server adjusts the state or behavior of the digital human model according to the received instructions. Use UnityProfiler to analyze the performance bottleneck of the scene and implement scene optimization measures. Scene optimization measures include eliminating objects that are not in the field of view and using Shader to reduce the number of dynamic light sources. Scene performance bottlenecks include objects that are out of sight and dynamic light sources that exceed the preset number in the scene. According to the position and role of the digital human in the scene, configure and adjust the Camera parameters in Unity, including the perspective and field of view, to obtain the best rendering perspective and visual effects. If the rendering content contains pre-recorded video, use the VideoPlayer component in Unity to control the video playback, and send the pre-recorded video to the client as part of the rendering stream synchronously. Encrypt the video stream data through the DTLS encryption protocol, test the encryption mechanism, determine whether the encryption affects the video stream transmission and rendering efficiency, and adjust the encryption algorithm and configuration accordingly. Integrate the real-time monitoring tool Prometheus into Unity to monitor the performance of the rendering stream in real time, including latency and frame loss. According to the monitoring results, adjust the encoding settings or network configuration to optimize the rendering stream quality.

[0040] For example, real-time rendering of digital people and dynamic scenes is provided for mobile phone users. The average delay of the network environment is 100ms, the bandwidth is 50Mbps, the packet loss rate is 1%, and the mainstream browsers of user devices include Chrome90 and Firefox88. The GPU performance of the rendering server is Teraflops12TFLOPs, and the CPU is 8 cores at 3.6GHz. The minimum standard for real-time rendering of high-complexity scenes is not met. Comparing H.264 and H.265, it is found that H.265 provides better visual quality under the same network conditions, although the bandwidth consumption is reduced by 20%, so H.265 is selected as the video encoder. Using Wireshark to monitor real-time data shows that the network delay occasionally rises to 150ms, and the resolution below 720p is enabled to optimize the RenderStreaming transmission quality. Based on the performance evaluation of the GPU and CPU, it is decided to upgrade the GPU to 24TFLOPs to meet the needs of real-time rendering. After testing, both Chrome90 and Firefox88 support the latest WebRTC features, and a compatibility solution based on the latest standards is selected. The designed WebSocket protocol successfully reduces the user input delay to less than 50ms. UnityProfiler pointed out that too many dynamic light sources are a performance bottleneck. By reducing it to only one main light source, the rendering frame rate is increased from 30fps to 60fps. The scene performance bottlenecks include objects in the scene that are beyond the line of sight and dynamic light sources that exceed the preset number. Use VideoPlayer to play pre-recorded videos to synchronize with the scene without obvious delay. The DTLS protocol test shows that the delay introduced by the encryption process does not exceed 10ms, and the encryption configuration is confirmed to have no impact on rendering efficiency. The integrated Prometheus monitoring tool shows that the average delay is within 65ms, the frame loss rate is less than 0.5%, and smooth transmission is maintained by automatically adjusting the encoding quality when network conditions fluctuate.

[0041] S102, deploying a distributed resource management module to monitor the load data of each computing node of the rendering server in real time, determine whether it is a faulty node, and reallocate the rendering tasks corresponding to the faulty node to non-faulty nodes, where the rendering tasks include digital human rendering and scene rendering.

[0042] By deploying the distributed resource management module Collectd, the load data of each computing node on the rendering server is obtained. The load data includes CPU utilization, memory usage, and network bandwidth. The obtained load data is sent to the central monitoring server, and the message middleware Kafka is used for real-time data transmission. The monitoring tool Prometheus is used to process and store the obtained load data to obtain the real-time load of each computing node. The data visualization tool Grafana is used to convert the processed performance data into charts and dashboards to monitor the status of the computing nodes of the rendering server in real time. If the monitoring data shows that the load data of any node exceeds the predetermined load threshold, the alarm and notification mechanism is triggered, and the node is judged to be a faulty node. The resource management and scheduling framework Kubernetes is used to dynamically allocate computing node resources according to real-time monitoring data. Cluster status management information is obtained to monitor the health status of each node in the cluster, and the automation tool Ansible is used to perform configuration management and maintenance tasks. If a computing node failure is detected, the fault-tolerance mechanism is activated to reallocate rendering tasks on the node to non-faulty nodes. Rendering tasks include digital human rendering and scene rendering. Log files generated by the rendering server are obtained through Logstash, and historical performance data and event logs are regularly analyzed to optimize resource scheduling strategies and improve fault recovery processes.

[0043] For example, there is a rendering server cluster consisting of 10 computing nodes. Each node is equipped with the same hardware configuration, with a 16-core Intel Xeon CPU, an NVIDIA RTX 2080 Ti GPU, 64 GB memory, and a network bandwidth of 1 Gbps. Deploy Collectd on each computing node and configure Collectd to collect data on CPU utilization, memory usage, and network bandwidth every 30 seconds. Transmit the data collected by Collectd to the central monitoring server through Kafka. The throughput of Kafka is configured to 5000 messages / sec, which is enough to handle the data streams from 10 nodes. Use Prometheus to process and store the received data. Configure Prometheus to pull data every minute to monitor the node load in real time. Use Grafana to create a dashboard that displays real-time charts of CPU utilization, memory usage, and network bandwidth. Set the thresholds for CPU utilization to 85%, memory usage to 80%, and network bandwidth usage to 75%. If the CPU utilization of a node exceeds 85% continuously, Grafana triggers an alarm and notifies the system administrator via email and SMS. When the CPU utilization of node 3 reaches 90%, an alarm is triggered immediately, and node 3 is judged to be a faulty node. Based on the real-time monitoring data provided by Prometheus, Kubernetes dynamically adjusts the rendering tasks on node 3 with higher load, and migrates some rendering tasks to nodes with lower load. The rendering tasks include digital human rendering and scene rendering. Using Ansible to regularly check the health status of each node, including the hardware status and system service status, it is found that the GPU temperature of node #7 is continuously higher than 80℃, which may require maintenance. When it is detected that node 5 cannot work due to hardware failure, Kubernetes automatically reallocates the rendering tasks on node 5 to non-faulty nodes to ensure that the rendering tasks are not affected. Through Logstash to obtain the log files generated by the rendering server, analyze the historical performance data and event logs, and find that the frequent restarts of node 2 are caused by software errors of memory leakage. After updating the software, the stability of node 2 is improved.

[0044] S103. According to the real-time monitoring data of each computing node of the rendering server, a performance evaluation is performed on each computing node of the rendering server, and a performance index of each computing node of the rendering server is determined by combining multi-dimensional performance indicators of each computing node, which include processing speed, stability and error rate.

[0045] The key performance indicators of each computing node of the rendering server are regularly obtained through the real-time monitoring tool Prometheus, including data on CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput. If the CPU or GPU usage in the obtained data continues to reach the preset peak value, it is determined whether there is uneven task distribution or hardware performance limitation. The statistical data of task completion time are used to evaluate the speed of task processing of each node of the rendering server. For nodes with processing time longer than the preset time threshold, the hardware and server configuration are checked to determine the cause of the delay. The load data of each computing node of the rendering server are used to determine the stability of each computing node of the rendering server when the load is higher than the preset load threshold. If the load of the computing node continues to be higher than the preset load threshold, it is determined that the node is at risk of overload. The log files generated by the rendering server are obtained through Logstash, and the data on the number of hardware failures, server crashes, and abnormal program exits are obtained from the log files to evaluate the stability of each node. The number of rendering errors and log error information are obtained from the log files generated by the rendering server, and the error rate of the node is calculated based on the number of rendering errors and log error information. If the error rate of the node increases, it is inferred that there is a problem with the hardware or software configuration. According to the multi-dimensional performance indicators, including processing speed, stability and error rate, weights are assigned to each multi-dimensional performance indicator, and the performance index calculation formula is adopted, PI=(ω v ×S v )+(ω s ×S s )+(ω e ×S e ), calculate the performance index of each computing node, where PI represents the performance index, ω v Represents the weight of processing speed, S v represents the score of processing speed, ω s represents the weight of stability, S s represents the stability score, ω e represents the weight of the error rate, S e The score represents the error rate. According to the calculated performance index, the computing nodes are sorted, and the nodes with performance lower than the preset performance threshold are identified, and the resource scheduling measures are adjusted.

[0046] For example, there is a rendering server cluster consisting of 5 computing nodes, each with different hardware configurations and workloads. Using Prometheus to monitor performance indicators, the performance indicator data collected by the real-time monitoring rendering server within a week is: Node 1 CPU usage 85%, GPU usage 90%, Node 2 CPU usage 60%, GPU usage 65%, Node 3 CPU usage 95%, GPU usage 98%, Node 4 CPU usage 50%, GPU usage 55%, Node 5 CPU usage 80%, GPU usage 85%. According to the data, the CPU and GPU usage of Node 3 is continuously higher than 90%, indicating that there may be uneven task distribution or hardware performance limitations. The average time for the nodes to process a rendering task is 30 minutes for Node 1, 20 minutes for Node 2, 40 minutes for Node 3, 15 minutes for Node 4, and 25 minutes for Node 5. If the preset time threshold is 25 minutes, the processing time of Node 3 and Node 1 is longer, and the hardware and rendering server configuration need to be checked. If the load threshold is 80%, the load of node 3 continues to exceed this threshold, and there is a risk of overload. The log file generated by the rendering server is obtained through Logstash. The log shows that the number of hardware failures and server crashes in a week are: 1 hardware failure in node 1 and 2 server crashes in node 3. Node 3 shows low stability. The number of rendering errors and log error information are obtained through the log file generated by the rendering server. The error rate of node 3 is 5%, and the error rates of other nodes are all less than 2%. According to multi-dimensional performance indicators, including processing speed, stability and error rate, weights are assigned to each multi-dimensional performance indicator, and the performance index calculation formula is used, PI = (ω v ×S v )+(ω s ×S s )+(ω e ×S e ), calculate the performance index of each computing node, where PI represents the performance index, ω v represents the weight of processing speed, S v represents the score of processing speed, ω s represents the weight of stability, S s represents the stability score, ω e represents the weight of the error rate, S e represents the score of error rate. If the processing speed, stability and error rate are 60, 70 and 80 respectively, and the weights of processing speed, stability and error rate are 0.5, 0.3 and 0.2 respectively, the performance index of node 3 is calculated as

[0047] PI = (0.5 × 60) + (0.3 × 70) + (0.2 × 80) = 67. The performance indexes of other nodes are calculated similarly. Based on the performance index, it is found that the performance index of node 3 is the lowest. It is decided to reduce the task allocation of node 3 and increase the task allocation of node 4 and node 2 to optimize the overall rendering efficiency.

[0048] S104. According to the deadline of the task in the video ringback tone rendering task, the historical digital human rendering data, and the scene rendering data, the urgency, complexity, and estimated rendering time of each rendering task are determined, and a task priority model is established to divide the rendering tasks into critical rendering tasks and non-critical rendering tasks. According to the performance index of each computing node of the rendering server and the priority of each rendering task, a node allocation score evaluation formula is constructed to generate a rendering task allocation strategy.

[0049] Get the rendering task schedule, determine the deadline for the video ringtone rendering task, and if the difference between the deadline and the current time is lower than the preset time threshold, mark the urgency of the rendering task as high. Adjust the urgency rating according to the customer's requirements for task completion time, and increase the urgency of rendering tasks with time requirements higher than the preset time threshold. Use database query to obtain historical digital human rendering data and scene rendering data. Digital human rendering data includes digital human model detail level, digital human texture resolution, digital human material complexity, and digital human dynamic effect requirements. Scene rendering data includes scene size and scene environment effects. Use decision tree algorithm to train the model and evaluate the complexity of the current rendering task. The complexity includes high, medium and low. Get hardware capability parameters, combine with rendering setting information, and use Blender simulation to estimate the expected rendering time of each rendering task. Hardware capability parameters include CPU and GPU performance indicators, and rendering setting information includes resolution and sampling rate. Get the availability of current hardware resources and human resources. If hardware resources and human resources are tight, for tasks with complexity higher than the preset complexity threshold and estimated rendering time longer than the preset rendering time threshold, increase the priority of the rendering task. Obtain the dependencies between rendering tasks. If the current rendering task is a prerequisite for multiple subsequent rendering tasks, increase the priority of the rendering task. According to the urgency, complexity, estimated rendering time and dependencies of the tasks, use the recurrent neural network algorithm to train the model, establish a task priority model, and perform a comprehensive score on each rendering task. According to the comprehensive score results, sort all rendering tasks and divide them into critical rendering tasks and non-critical rendering tasks according to the score. Tasks with a comprehensive score higher than the preset score threshold are classified as critical rendering tasks, and tasks with a comprehensive score lower than the preset score threshold are classified as non-critical rendering tasks. Obtain the performance index, load condition and rendering task priority of each computing node of the rendering server, construct a node allocation score evaluation formula, node allocation score = λ×performance index-μ×load condition, allocate the optimal computing node to each rendering task, and generate a rendering task allocation strategy, where λ and μ are weight coefficients used to balance the impact of performance and load. According to the generated rendering task allocation strategy, the rendering tasks are allocated to the selected computing nodes, and critical rendering tasks are given priority;

[0050] For example, there is a rendering server cluster that is used to perform various rendering tasks. Currently, there is a new rendering project that needs to be evaluated and resourced. The deadline for rendering task A is 10 days later, while the preset time threshold is 7 days, so the urgency of task A is marked as low. The deadline for rendering task B is 5 days later, which is lower than the preset time threshold, so the urgency of task B is marked as high. The customer pointed out that task B needs to be completed within 4 days, so the urgency rating of task B is further increased. The database query shows that the model detail level is complex animation with high dynamic performance, the texture resolution is high-definition 4096x4096, the material complexity is complex materials with multi-level reflection, transparency and lighting effects, and the dynamic effect requirements are real-time physical calculations and dynamic interactions of particle systems. The scene size is a large open world, and the scene environment effect is a complex dynamic weather system and real-time ray tracing effects. The decision tree algorithm is used to train the model with the above collected data to evaluate the complexity of the current rendering task. Through algorithmic analysis, tasks are divided into three complexity levels: high, medium, and low. A task is evaluated as high complexity if it includes a digital human model with a high level of detail, 4K texture resolution, high material complexity, and complex dynamic effect requirements. Conversely, if a task involves a model with a low level of detail, basic texture resolution and materials, and no dynamic effects are required, it is evaluated as low complexity. Using Blender and hardware performance parameters, with a 16-core CPU and an RTX2080Ti GPU, the expected rendering time for Task A is 72 hours and for Task B is 24 hours. Currently, hardware resources are tight and human resources are sufficient. Therefore, for Task A, which has high complexity and a long estimated rendering time, its priority is increased. Task B is a prerequisite for multiple subsequent rendering tasks, so the priority of Task B is further increased. Collect relevant data of rendering tasks, including Task A with an urgency of 1, a complexity of 2, an estimated rendering time of 3 hours, and a dependency of 0; Task B with an urgency of 0, a complexity of 1, an estimated rendering time of 2 hours, and a dependency of 1, and depends on Task A. Use the recurrent neural network algorithm to train the model and establish a task priority model. The comprehensive score of Task A is 85 and that of Task B is 50. The preset score threshold is 80, so Task A is classified as a critical rendering task and Task B is a non-critical rendering task. According to the score and priority, the highest performance hardware resources, such as RTX2080TiGPU, and priority human support are allocated to the critical rendering task A to ensure timely completion. Non-critical Task B is processed after Task A is completed. If there is a rendering server cluster consisting of 3 computing nodes, the performance index PI, current load status, and rendering tasks to be processed of each node are: PI of Node 1 = 90, current load = 60%, PI of Node 2 = 80, current load = 40%, and PI of Node 3 = 70, current load = 50%.There are two rendering tasks, Task A and Task B, using the node allocation score evaluation formula, where λ = 0.5, μ = 0.5. The node allocation scores are calculated as follows: Node 1 score = 0.5×90-0.5×60 = 15, Node 2 score = 0.5×80-0.5×40 = 20, and Node 3 score = 0.5×70-0.5×50 = 10. Based on the allocation scores, Task A is assigned to Node 2 with the highest score, while Task B is assigned to Node 1 with the second highest score. During the execution of the tasks, it was monitored that the response time of Node 1 increased from an average of 500ms to 800ms, and the playback smoothness of Node 2 decreased from 30 frames per second to 20 frames per second.

[0051] S105. Adjust the priority of non-critical rendering tasks based on user experience indicators, which include response time and playback smoothness. According to the user attributes of the callee, determine the preference of digital human rendering and scene rendering in the callee's video ringback tone, and adjust digital human rendering parameters and scene rendering parameters based on the preference of digital human rendering and scene rendering in the callee's video ringback tone.

[0052] Monitor user experience indicator data in real time, including response time and playback smoothness. Based on the obtained user experience indicator data, use the support vector machine algorithm to train the model to determine whether there is a downward trend in user experience quality. If there is a downward trend in user experience quality, implement a dynamic priority adjustment mechanism to dynamically reduce the priority of non-critical rendering tasks being processed based on real-time user experience indicators, and give priority to rendering tasks that affect user experience. Based on historical user experience data, use the decision tree algorithm to train the model, predict changes in user experience indicators, and adjust task priorities in advance before user experience indicators decrease. During the execution of rendering tasks, continuously monitor the execution status and user experience indicators of each rendering task, and adjust the rendering task allocation strategy based on real-time data; obtain the device performance parameters of the call object, including the number of CPU cores, memory size, and GPU type through the device performance detection library navigator.hardwareConcurrency. Set a performance baseline based on the device performance parameters to determine whether the device can carry high, medium, and low levels of rendering complexity. Use the network monitoring tool SpeedtestAPI to obtain the network bandwidth and connection speed data of both parties in the call. If the network bandwidth of either party is lower than the predetermined threshold, reduce the resolution and frame rate of the video call. Add user preference options to the application settings, including video resolution preferences and quality or speed priority options, and use user profiles stored locally or in the cloud to obtain user preferences. According to user preferences and environmental sensor data, determine the brightness and background noise level of the environment in which the caller is located, and adjust the brightness and volume settings of the video ringback tone. According to the user attributes and historical usage data of the caller, use the decision tree algorithm to train the model and predict the video ringback tone style preferences of the caller. User attributes include age, gender, and hobbies. According to the video ringback tone style preferences of the caller, adjust the rendering style of the digital human and scene; according to the device performance, network conditions, video ringback tone style preferences, and environmental data of the caller, use the recurrent neural network algorithm to train the model and obtain the optimal digital human rendering parameter settings and scene rendering parameter settings for the call. The digital human rendering parameters include the digital human model detail level, digital human texture resolution, digital human material complexity, and digital human dynamic effect requirements. The scene rendering parameters include scene size and scene environment effects.

[0053] For example, the user experience indicator data is obtained, and it is found that the average node response time increased from 250ms to 350ms in the past week, and the playback fluency increased from 0.5% frame loss rate to 2% frame loss rate in the past week. The daily response time and playback fluency in the past week are used as features to train the model with the support vector machine algorithm. After the model is trained, the current response time of 350ms and the frame loss rate of 2% are input, and the model predicts that the quality of user experience will decline in the future. Based on the analysis of the support vector machine algorithm, the priority of non-critical rendering tasks on node 1 is dynamically lowered to ensure that critical task A can be processed first and alleviate the downward trend of user experience quality. The decision tree algorithm is used to analyze historical user experience data and predict that the response time may continue to rise in the next hour. Based on this, the management system adjusts the priority of task B on node 2 in advance to prevent further decline in user experience. During the execution of the rendering task, the performance indicators and user experience indicators of the node are continuously monitored. Based on real-time data, some small rendering tasks that can significantly improve the user experience are processed in advance, the response time is reduced to below 500ms, and the playback fluency is restored to more than 90%. There are two callers, A and B. They use devices with different performances to make video calls, and both like high-quality video ringback tone services. A uses a high-performance device with an 8-core CPU, 16GB memory, and a high-end GPU, such as NVIDIARTX3060. Alice's network connection speed is 100Mbps download speed and 50Mbps upload speed. She prefers high-resolution video ringback tone, and the environment where she is located is dark. B uses a mid-range performance device with a 4-core CPU, 8GB memory, and a medium-level GPU, such as NVIDIAGTX1660. B's network connection speed is 20Mbps download speed and 10Mbps upload speed. B has no special preference, but the background noise in his environment is large. For A's device, it can carry high-complexity rendering tasks, and for B's device, it is recommended to carry medium-complexity rendering tasks. A's network environment supports high-quality video transmission, while B's network environment can only support medium-quality video. Through the application settings, it is known that A prefers high-resolution video ringback tone, and B has no specific preference, but the background noise is large and needs noise reduction. Since the light in A's environment is low, the brightness of the video ringback tone is automatically enhanced, and noise reduction is started for B to improve the call quality. The model predicts that the user may prefer dynamic and high-energy video ringback tone, so a ringback tone with fast-paced music and dynamic visual effects is automatically set for the upcoming call. The user attributes of the call object are obtained. A is 23 years old, male, and likes games, music, and sports. B is 30 years old, female, and likes shopping and watching TV series. The decision tree algorithm is used for model training. It is predicted that A tends to prefer a dynamic and highly dynamic digital human rendering style, while B has no obvious preference for the video ringback tone style. User attributes include age, gender, and hobbies.If A is using a mid-range smartphone with an 8-core CPU, 6GB RAM and a medium-performance GPU, such as Adreno640, the network bandwidth in A's area is 20Mbps, and the latency is about 50ms. A indicates through the application settings that he likes high-definition videos and provides preference information for quiet forest scenes. In addition, the environmental data shows that the user's area is at night and the indoor light is dim. Based on the device performance, network conditions, video ringtone style preferences, and environmental data of the call object, the recurrent neural network model determines that the device can carry medium to high levels of rendering complexity. Therefore, the target resolution of the video ringtone is set to 1080p and the frame rate is 30fps, ensuring high visual quality. The user's preferred forest scene and high-definition video prompt model increase the level of detail of the digital human model, such as using high-resolution textures and more complex materials. At the same time, the scene is set to a larger forest environment, including a variety of trees and vegetation, as well as dynamic effects such as breeze blowing leaves and randomly flying butterflies to increase the vitality of the scene. Considering the dim indoor light, the model adjusted the scene's lighting parameters, adding virtual moonlight and soft ambient light to ensure that users can clearly see the digital human and scene details without feeling visual discomfort. Finally, the optimal rendering parameters output by the recurrent neural network model were set to: Digital human rendering parameters include high model detail level, 2048x2048 texture resolution, medium material complexity, including reflection and transparency processing, and dynamic effects including simple facial expressions and gesture animations. Scene rendering parameters include setting the scene size to a large forest, and environmental effects including dynamic wind effects and slight fog effects to increase the sense of depth and realism.

[0054] S106. Monitor network bandwidth utilization and latency through Nagios, perform data pre-fetching and data synchronization based on the monitoring results, use LRU and FIFO to manage the cache servers corresponding to each computing node, and dynamically allocate rendering tasks based on the status of the cache server. Continuously optimize the network environment and cache server configuration based on the cache server load and performance indicators.

[0055] Through the network monitoring tool Nagios, the bandwidth utilization and latency of the current network are continuously detected. If changes in the network environment are detected, the pre-fetch mechanism is triggered. Changes in the network environment include bandwidth utilization increases greater than the preset increase threshold or delay increases greater than the preset increase threshold. According to the pre-fetch trigger conditions, a data pre-fetch strategy is adopted to obtain historical access patterns and user behavior data related to the task, and a decision tree algorithm is used for model training to determine the resource data that should be pre-loaded into the cache. If it is determined that the resource data needs to be pre-fetched, a timestamp mechanism is applied to synchronize the source data and the data in the cache. After completing the data pre-fetch, the data expiration strategy in the cache strategy is used to manage the data in the cache. The data expiration strategy includes the least used LRU and the first-in-first-out FIFO. Based on the node allocation score evaluation formula, the rendering task is dynamically allocated according to the resource status and task requirements of the cache server. If it is detected that the resource load of the specified server is higher than the preset second load threshold, the rendering task to be allocated is reallocated to the server whose load is lower than the preset third load threshold. Evaluate the performance of the cache server based on the hit rate and service quality indicators in the cache server's performance indicators. If the cache server's performance does not meet the standards, adjust the cache size or increase the number of servers to expand the cache server's performance level. Evaluate the cache server's availability indicators based on the proportion of time the cache server is operating normally, and obtain response time data, where the response time is the time from the user initiating a request to the cache server responding. Continuously evaluate the user's user experience quality during the rendering task execution process through the response time and cache server availability indicators. If user experience problems are detected, optimize the network environment and cache server configuration. Apply a fault recovery strategy. When a cache server fails, switch to a backup server or reallocate resources. The fault recovery strategy includes a backup mechanism and a fault tolerance mechanism.

[0056] For example, an online rendering service is being operated, which relies on a distributed cache system to improve the access speed of rendering data and reduce network latency. Use Nagios to monitor the current network status, set the preset increase threshold of bandwidth utilization to 50% and the preset increase threshold of latency to 100ms. In a monitoring cycle, the Nagios report shows that the bandwidth utilization has increased from 30% to 85%, and the latency has increased from 50ms to 200ms, exceeding the preset thresholds, triggering the data pre-fetching mechanism. Based on user behavior data and historical access patterns, the decision tree algorithm is used to analyze the most likely rendering resources to be requested. The analysis shows that from 7 to 9 p.m., the most frequent user requests are for the rendering data of the urban landscape. This part of the rendering data of the urban landscape is preloaded into the cache, and the data is synchronized with the source server. The timestamp mechanism is used to manage the consistency of the data. The cache policy is set to use the least recently used LRU strategy to manage expired data to ensure efficient use of cache space. In a rendering project, there are two key servers, including cache servers A and B. If the average load of cache server A has reached 80% in the past 30 minutes, and the performance index is 90, according to the node allocation score evaluation formula, the score is 0.5×90-0.5×80=5, then the score = 5. Since server B has a higher score of 10, it is necessary to migrate some rendering tasks from A to B to reduce the load of A. The average hit rate of the cache server is maintained at 95%, indicating good performance. The response time data shows that the average response time has been reduced from 300ms before request optimization to 150ms, meeting the goal of user experience optimization. In subsequent monitoring, if it is found that the response time of a cache server continues to exceed 200ms, the network environment may be adjusted or the number of cache servers may be increased to ensure service quality. Once Nagios monitors that a cache server cannot respond due to hardware failure, the backup mechanism is immediately enabled, the traffic is switched to the backup server, and the technical team is notified to repair the fault.

[0057] S107, regularly checking the health status of each computing node of the rendering server. If the response time of the computing node exceeds a preset threshold, it is identified as an unhealthy node and triggers task reallocation, including making decisions based on performance evaluation data and task classification data, and migrating the tasks of the unhealthy node to other healthy nodes with matching performance.

[0058] Implement a scheduled task mechanism, set the rendering task detection interval, detect the health detection indicators of each computing node of the rendering server, and evaluate the health status of each computing node. The health detection indicators include CPU usage and memory usage. Regularly obtain the response time data of each node, and use the monitoring tool Prometheus combined with Grafana to determine whether the node is within the preset response time range. For nodes with abnormal response time or other abnormal health detection indicators, perform retry detection. If the number of consecutive detection failures is greater than the preset threshold, the node is determined to be an unhealthy node. Obtain the error log of the unhealthy node, use the abnormal information to assist in determining the fault type of the node, and determine the corresponding fault repair measures. Obtain the task information related to the unhealthy node from the task queue and the mapping information between the task and the computing node, and evaluate the current status of the task. For each task to be migrated, calculate the performance matching degree between the migration task and each normally operating node based on the resource requirements of the migration task and the resource supply capacity of the node. The resource requirements include CPU intensive and memory intensive. Based on the urgency of the task, resource requirements, and user needs, the decision tree algorithm is used for model training to evaluate the priorities of all tasks to be migrated, including high priority, medium priority, and low priority. The migration strategy is defined based on the priority of the task and the performance matching metric. The migration strategy includes determining the order of migration tasks and selecting the best target node. High-priority migration tasks are migrated first and migrated to the node with the highest performance match. According to the migration strategy, the tasks of unhealthy nodes are migrated to other healthy nodes with matching performance.

[0059] For example, there is a rendering server cluster consisting of 4 computing nodes. The configuration and role of each node are: Node 1 has 16 CPU cores and 32 GB memory; Node 2 has 16 CPU cores and 64 GB memory, which is an unhealthy node; Node 3 has 32 CPU cores and 64 GB memory; Node 4 has 8 CPU cores and 16 GB memory. Implement a scheduled task mechanism, use Prometheus to detect the CPU usage and memory usage of the nodes every 5 minutes, and use Grafana to obtain the average response time of each node. The preset response time range is within 200 ms. In one test, it was found that the CPU usage of Node 2 was continuously 100% and the response time exceeded 500 ms. After 3 consecutive tests, the preset number of thresholds was 2 times, and Node 2 was determined to be an unhealthy node. Obtain the error log of Node 2 to display the hardware failure prompt information, and further analysis determined that the CPU hardware failed. From the task queue, we identify three rendering tasks processed by node 2. Task A has a CPU-intensive resource requirement, high urgency, and high user demand. Task B has a memory-intensive resource requirement, medium urgency, and low user demand. Task C has a balanced resource requirement, low urgency, and medium user demand. Task A is CPU-intensive, so it has the best performance match with the 32-core CPU of node 3. Task B is memory-intensive, so it also has the best performance match with the 64GB memory of node 3. Task C has a certain degree of match with both node 1 and node 4, but node 1 has more abundant resources. Using the decision tree algorithm, we evaluate the priorities of tasks A, B, and C as high, medium, and low. Based on the priority and performance match, we define a migration strategy. Task A is migrated to node 3 first, followed by task B and task C. According to the migration strategy, we automatically start migrating tasks A, B, and C from the failed node 2 to the target node. After the migration is completed, we continuously monitor the execution status and response time of the tasks on the new node to ensure that the migration is successful without performance loss.

[0060] S108. An audio and video call link is established between the rendering server and the user terminal through the Coturn server, and the rendered digital human and scene are transmitted through the audio and video call link after the rendering task is completed.

[0061] According to the network environment of the rendering server and the user terminal, the STUN service is configured through the Coturn server to obtain the public IP and port information of the rendering server and the user terminal, and determine whether there are NAT or firewall restrictions. If the user terminal and the rendering server are both behind NAT, the TURN service of the Coturn server is used to establish a relay to transmit data between devices behind different NATs. The security protocol DTLS is used to encrypt the audio and video data streams to obtain an audio and video transmission solution that protects the communication link. The WebRTC protocol is used to establish a real-time audio and video call link between the rendering server and the user terminal to evaluate the bandwidth and delay of the link, and optimize the quality of audio and video transmission according to the bandwidth and delay of the link. According to the transmission quality of the real-time call link, the audio and video properties are adjusted, including reducing the resolution or frame rate, to obtain the best call effect adapted to the current network environment. The processing capacity of the rendering server is judged through the performance parameters of the rendering server, and the complexity of the rendering task is determined based on the processing capacity of the rendering server. The performance parameters of the rendering server include CPU performance and GPU performance. The device type and operating system information of the user terminal are obtained, the compatibility and display effect of the rendering content on the user terminal are evaluated, and the optimal rendering parameters for the user device are determined. Through the WebRTC session control protocol SessionDescriptionProtocol, a session management mechanism for audio and video calls is established to synchronously transmit the rendering tasks of the digital human model and the scene on the call link. After the rendering task is completed, the rendered digital human and scene are transmitted to the user terminal through the established secure and encrypted audio and video call link to achieve real-time audio and video interaction.

[0062] For example, the STUN service is configured using the Coturn server. The public IP of the rendering server is 203.0.113.1, port 3478, and the public IP of the user terminal is 198.51.100.1, port 3478. It is found that the user terminal is behind NAT and the rendering server is also behind different NATs. Since both the user terminal and the rendering server are behind NAT, Coturn's TURN service is enabled to establish a relay between the two. The IP of the TURN server is 203.0.113.2, port 3478. The DTLS protocol is used to encrypt the audio and video data streams. The DTLS protocol version 1.2 ensures the security of communication. Through the WebRTC protocol, a real-time audio and video call link is established between the rendering server and the user terminal. The link test shows that the bandwidth is about 10Mbps and the average delay is 100ms. Based on the current bandwidth and delay, the initial audio and video resolution is set to 1920x1080 and the frame rate is 30fps. Due to bandwidth fluctuations, the resolution is dynamically adjusted to 1280x720, and the frame rate remains unchanged to adapt to network conditions. The rendering server is equipped with an Intel Xeon E5 CPU, 32 cores, and an NVIDIA RTX 2080 Ti GPU. Based on the server performance, it is determined that the rendering task can support high-complexity model rendering. The user device is an iPhone 12, and the operating system is iOS 14. The optimal rendering parameters for the device are determined to be a resolution of 720x1280 to ensure good display effects and performance. The audio and video format, encryption method and other parameters are negotiated through Session Description Protocol to ensure smooth audio and video calls. After the rendering of the digital human model and scene is completed, the rendered digital human model and scene are transmitted to the user terminal in real time through the established encrypted audio and video call link.

[0063] S109: Based on the historical performance data of each computing node of the rendering server, a node performance prediction model is constructed to obtain the load and performance changes of each computing node of the rendering server in a future preset time period, determine the performance bottleneck of each computing node, and perform rendering task reallocation in advance.

[0064] Obtain key performance indicator data of each computing node of the rendering server, including data on CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput. Based on historical rendering task data, identify the demand for computing resources and node resource consumption patterns of different types of rendering tasks. Different types of rendering tasks include CPU-intensive, GPU-intensive, and memory-intensive tasks. Organize the obtained key performance indicator data in timestamp sequence, use the ARIMA algorithm for model training, predict the key performance indicator values ​​in the future preset time period, and determine the periodic fluctuations and time dependence of node performance based on the predicted trend line, seasonal component, and residual. Based on historical task execution data, use the method of calculating the Pearson correlation coefficient to determine the correlation between task execution mode and node performance. Historical task execution data includes task duration, resource consumption, and queue waiting time. Obtain system event data from system maintenance logs, hardware replacement records, and abnormal event logs. For each obtained system event, associate the computing node performance indicator data within the corresponding time period. According to the performance indicator changes before and after the system event occurs, identify the impact of the system event on the computing node performance and task execution. The system event data includes the date and time of maintenance, maintenance activity description, hardware replacement parts and reasons, and the time, type, and impact range of abnormal events. Obtain key hardware configuration information of each computing node of the rendering server, including the number of CPU cores, CPU model, GPU model and computing power, memory capacity, disk type, disk capacity, and network interface type. Associate the obtained hardware configuration information with the performance indicator data of the node, use the variance analysis method to determine the correlation between the hardware configuration parameters and the node performance indicators, and identify the hardware configuration factors that affect the node performance to a greater extent than the preset threshold. Based on the hardware configuration, performance indicators, task characteristics, and system event log data, use the random forest algorithm to train the model, build a node performance prediction model, and predict the performance changes of the computing node in the future preset time period. According to the prediction results of the node performance prediction model, identify potential performance bottleneck areas, where the performance bottleneck is the node with degraded performance. Based on the prediction results of the node performance prediction model and the performance bottleneck prediction, the distribution of rendering tasks is adjusted in advance. If a node is predicted to face a performance bottleneck, the rendering tasks are reallocated in advance.

[0065] For example, there is a rendering server cluster consisting of 3 computing nodes. The key performance indicator data obtained is the average performance indicator data of each node in the past week. The CPU usage of node 1 is 70%, the GPU usage is 60%, the memory usage is 80%, the disk I / O throughput is 200MB / s, and the network throughput is 100MB / s. The CPU usage of node 2 is 80%, the GPU usage is 70%, the memory usage is 90%, the disk I / O throughput is 250MB / s, and the network throughput is 120MB / s. The CPU usage of node 3 is 60%, the GPU usage is 50%, the memory usage is 75%, the disk I / O throughput is 180MB / s, and the network throughput is 90MB / s. By analyzing the historical rendering task data, it is identified that most rendering tasks are GPU intensive. The CPU usage of node 1 was trained and predicted using the ARIMA algorithm. The model training showed that the CPU usage of node 1 had a clear periodic fluctuation pattern of weekend lows and weekday peaks. The trend line of CPU usage predicted for the next month showed that with the launch of a new project, the usage is expected to rise to an average of 75% and maintain the pattern of weekend lows and weekday peaks. By calculating the Pearson correlation coefficient, it was found that the task duration was positively correlated with the CPU usage with a coefficient of 0.85, indicating that an increase in CPU usage will lead to an increase in task execution time. After a hardware replacement activity GPU replacement on node 2, the GPU usage dropped from 70% to 50%, showing that the hardware replacement had a significant positive impact on performance. Variance analysis showed that there was a significant correlation between GPU computing power and GPU usage, especially in GPU-intensive rendering tasks. Get the hardware configuration. The hardware configuration of node 1 is CPU core number = 8, CPU model = XeonE5-2640v4, GPU model = NVIDIAGTX1080, memory capacity = 32GB, disk type = SSD, disk capacity = 1TB, network interface type = 10GbEthernet. Input the hardware configuration parameters, historical performance indicator data, task characteristics and system event log as features into the random forest model to build a node performance prediction model. The historical performance indicator data includes CPU usage and GPU usage. The system event log includes two hardware maintenances in the past three months. The node performance prediction model predicts that with the addition of new projects and the expected increase in rendering tasks, node 1 may face a CPU usage of more than 80% in the next month, suggesting a potential performance bottleneck. Based on the results of the node performance prediction model, it is decided to migrate some memory-intensive rendering tasks from node 1 to node 2 in advance to avoid the predicted performance bottleneck.

[0066] The above embodiment is only one of the preferred implementation modes of the present invention and should not be used to limit the protection scope of the present invention. Any changes or modifications that are made to the main design concept and spirit of the present invention and have no substantive significance, and the technical problems they solve are still consistent with the present invention, should be included in the protection scope of the present invention.

Claims

1. A method for constructing a communication network video ringback tone XR application, characterized in that: The method comprises: Based on the Render Streaming component of the Unity environment, the digital human and scene in the video ringback tone are rendered on the rendering server; Deploy a distributed resource management module to monitor the load data of each computing node of the rendering server in real time, determine whether it is a faulty node, and reallocate the rendering tasks corresponding to the faulty node to the non-faulty node. The rendering tasks include digital human rendering and scene rendering. Based on the real-time monitoring data of each computing node of the rendering server, the performance of each computing node of the rendering server is evaluated, and the performance index of each computing node of the rendering server is determined by combining the multi-dimensional performance indicators of each computing node, including processing speed, stability and error rate; According to the deadline of the task in the video ringtone rendering task, the historical digital human rendering data and the scene rendering data, and the urgency, complexity and estimated rendering time of each rendering task are determined, a task priority model is established, and the rendering tasks are divided into critical rendering tasks and non-critical rendering tasks. According to the performance index of each computing node of the rendering server and the priority of each rendering task, a node allocation score evaluation formula is constructed to generate a rendering task allocation strategy; Adjust the priority of non-critical rendering tasks based on user experience indicators, including response time and playback smoothness. According to the user attributes of the callee, determine the preference of digital human rendering and scene rendering in the callee's video ringback tone, and adjust the digital human rendering parameters and scene rendering parameters based on the preference of digital human rendering and scene rendering in the callee's video ringback tone. Monitor network bandwidth utilization and latency through Nagios, perform data pre-fetching and data synchronization based on monitoring results, use LRU and FIFO to manage cache servers corresponding to each computing node, dynamically allocate rendering tasks based on the status of the cache server, and continuously optimize the network environment and cache server configuration based on cache server load and performance indicators; Regularly check the health status of each computing node of the rendering server. If the response time of a computing node exceeds a preset threshold, it is identified as an unhealthy node and triggers task reallocation, including making decisions based on performance evaluation data and task classification data to migrate tasks of unhealthy nodes to other healthy nodes with matching performance. An audio and video call link is established between the rendering server and the user terminal through the Coturn server, and the rendered digital human and scene are transmitted by the audio and video call link after the rendering task is completed; Based on the historical performance data of each computing node of the rendering server, a node performance prediction model is constructed to obtain the load and performance changes of each computing node of the rendering server in the future preset time period, determine the performance bottleneck of each computing node and perform the redistribution of rendering tasks in advance; Establish a task priority model, divide rendering tasks into critical rendering tasks and non-critical rendering tasks, and construct a node allocation score evaluation formula based on the performance index of each computing node of the rendering server and the priority of each rendering task, and generate a rendering task allocation strategy, including: using a recurrent neural network algorithm to train the model according to the urgency, complexity, estimated rendering time and dependency of the task, establish a task priority model, and perform a comprehensive score on each rendering task; According to the comprehensive scoring results, all rendering tasks are sorted and divided into key rendering tasks and non-key rendering tasks according to the scores. The tasks with comprehensive scores higher than the preset scoring threshold are classified as key rendering tasks, and the tasks with comprehensive scores lower than the preset scoring threshold are classified as non-key rendering tasks. Obtain the performance index and load of each computing node of the rendering server and the rendering task priority, construct a node allocation score evaluation formula, node allocation score = λ×performance index-μ×load, allocate the best computing node for each rendering task, and generate a rendering task allocation strategy, where λ and μ are weight coefficients used to balance the impact of performance and load; According to the generated rendering task allocation strategy, the rendering tasks are allocated to the selected computing nodes, and key rendering tasks are prioritized; The priority of non-critical rendering tasks is adjusted based on user experience indicators, where the user experience indicators include response time and playback smoothness, including: real-time monitoring of user experience indicator data, including focusing on response time and playback smoothness; Based on the acquired user experience indicator data, the support vector machine algorithm is used to train the model to determine whether there is a downward trend in user experience quality; If there is a trend of declining user experience quality, a dynamic priority adjustment mechanism will be implemented to dynamically reduce the priority of non-critical rendering tasks being processed based on real-time user experience indicators, and give priority to rendering tasks that affect user experience.

2. The method according to claim 1, wherein: The Render Streaming component based on the unity environment renders the digital human and scene in the video ringback tone on the rendering server, including: By comparing coding efficiency and visual quality with bandwidth consumption, the performance of different encoders under the same network conditions is evaluated, the best matching encoder is selected, and the encoder configuration is updated regularly. Different encoders include H.264 or H.265; Use the network quality monitoring tool Wireshark to obtain real-time data on packet loss rate, latency, and bandwidth. If the real-time data shows that the network condition causes the quality of RenderStreaming to deteriorate, adjust the transmission strategy, including changing the encoding quality or enabling a resolution lower than the preset resolution threshold. Based on the complexity of the digital human model and scene in the video ringback tone, evaluate the current GPU performance and CPU performance of the rendering server to determine whether the preset minimum standards for real-time rendering of the digital human model and scene are met. If the server performance does not meet the minimum standards for real-time rendering of the digital human model and scene, perform hardware upgrades to improve the server's rendering capabilities. Select the best WebRTC compatibility solution based on the client browser's support for WebRTC, including browser version and platform features, as well as support for WebRTC core functions, and test the rendering effect and fluency on different browsers. WebRTC core functions include audio and video calls and data channels; Design a real-time communication protocol based on WebSocket to encode user input and transmit it to the rendering server in real time. The rendering server adjusts the state or behavior of the digital human model according to the received instructions. Use UnityProfiler to analyze scene performance bottlenecks and implement scene optimization measures, including removing objects that are out of view and using Shader to reduce the number of dynamic light sources. Scene performance bottlenecks include objects that are out of view and dynamic light sources that exceed the preset number in the scene. Configure and adjust the Camera parameters in Unity, including the viewing angle and field of view, according to the position and role of the digital human in the scene to obtain the best rendering perspective and visual effects; If the rendered content contains pre-recorded video, use the VideoPlayer component in Unity to control video playback, and send the pre-recorded video to the client as part of the rendering stream; Encrypt video stream data through the DTLS encryption protocol and test the encryption mechanism to determine whether encryption affects video stream transmission and rendering efficiency, and adjust the encryption algorithm and configuration accordingly; Integrate the real-time monitoring tool Prometheus into Unity to monitor the performance of the rendering stream in real time, including latency and frame loss. Based on the monitoring results, adjust the encoding settings or network configuration to optimize the rendering stream quality.

3. The method according to claim 1, wherein: The distributed resource management module is deployed to monitor the load data of each computing node of the rendering server in real time, determine whether it is a faulty node, and reallocate the rendering tasks corresponding to the faulty node to the non-faulty node. The rendering tasks include digital human rendering and scene rendering, including: By deploying the distributed resource management module Collectd, on each computing node of the rendering server, the load data of each computing node is obtained. The load data includes CPU utilization, memory usage, and network bandwidth. The acquired load data is sent to the central monitoring server, and the message middleware Kafka is used for real-time data transmission; Use the monitoring tool Prometheus to process and store the acquired load data and obtain the real-time load of each computing node; Use the data visualization tool Grafana to convert the processed performance data into charts and dashboards to monitor the status of the rendering server computing nodes in real time; If the monitoring data shows that the load data of any node exceeds the predetermined load threshold, the alarm and notification mechanism is triggered and the node is judged to be a faulty node; Adopt the resource management and scheduling framework Kubernetes to dynamically allocate computing node resources based on real-time monitoring data; Obtain cluster status management information, monitor the health of each node in the cluster, and use the automation tool Ansible to perform configuration management and maintenance tasks; If a computing node failure is detected, the fault tolerance mechanism is activated to reallocate rendering tasks on the node to non-faulty nodes. Rendering tasks include digital human rendering and scene rendering. Use Logstash to obtain log files generated by the rendering server, regularly analyze historical performance data and event logs, optimize resource scheduling strategies, and improve fault recovery processes.

4. The method according to claim 1, wherein: The method of performing performance evaluation on each computing node of the rendering server according to real-time monitoring data of each computing node of the rendering server and determining the performance index of each computing node of the rendering server in combination with multi-dimensional performance indicators of each computing node, wherein the multi-dimensional performance indicators include processing speed, stability and error rate, includes: Use the real-time monitoring tool Prometheus to regularly obtain key performance indicators of each computing node of the rendering server, including CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput data; If the CPU or GPU usage in the acquired data continues to reach the preset peak value, determine whether there is uneven task distribution or hardware performance limitation; Evaluate the speed at which each node of the rendering server processes tasks through statistical data on task completion time. For nodes whose processing time is longer than the preset time threshold, check the hardware and server configuration to determine the cause of the delay. Determine the stability of each computing node of the rendering server when the load is higher than a preset load threshold through the load data of each computing node of the rendering server. If the load of the computing node is continuously higher than the preset load threshold, it is determined that the node is at risk of overload. Obtain the log files generated by the rendering server through Logstash, obtain data on the number of hardware failures, server crashes, and program abnormal exits from the log files, and evaluate the stability of each node; Obtain the number of rendering errors and log error information from the log files generated by the rendering server, and calculate the node error rate based on the number of rendering errors and log error information. If the node error rate increases, it is inferred that there is a problem with the hardware or software configuration; According to the multi-dimensional performance indicators, including processing speed, stability and error rate, weights are assigned to each multi-dimensional performance indicator, and the performance index calculation formula is adopted. PI = (ωv × Sv) + (ωs × Ss) + (ωe × Se), calculates the performance index of each computing node, where PI represents the performance index, ωv represents the weight of the processing speed, Sv represents the score of the processing speed, ωs represents the weight of the stability, Ss represents the score of the stability, ωe represents the weight of the error rate, and Se represents the score of the error rate; According to the calculated performance index, the computing nodes are sorted, the nodes whose performance is lower than the preset performance threshold are identified, and the resource scheduling measures are adjusted.

5. The method according to claim 1, wherein: The method of determining the urgency, complexity and estimated rendering time of each rendering task according to the deadline of the task in the video ringback tone rendering task and the historical digital human rendering data and scene rendering data includes: Obtain the rendering task schedule, determine the deadline of the video ringback tone rendering task, and if the difference between the deadline and the current time is lower than a preset time threshold, mark the urgency of the rendering task as high; Adjust the urgency rating based on the customer's requirements for task completion time, and increase the urgency of rendering tasks with time requirements higher than the preset time threshold; Use database query to obtain historical digital human rendering data and scene rendering data. The digital human rendering data includes the digital human model detail level, digital human texture resolution, digital human material complexity, and digital human dynamic effect requirements. The scene rendering data includes scene size and scene environment effects. Use the decision tree algorithm to train the model and evaluate the complexity of the current rendering task. The complexity includes high, medium and low. Obtain hardware capability parameters, combined with rendering setting information, and use Blender simulation to estimate the expected rendering time of each rendering task. Hardware capability parameters include CPU and GPU performance indicators, and rendering setting information includes resolution and sampling rate. Obtain the availability of current hardware resources and human resources. If the hardware resources and human resources are tight, the priority of the rendering task is increased for the task whose complexity is higher than the preset complexity threshold and whose estimated rendering time is longer than the preset rendering time threshold. Get the dependency relationship between rendering tasks. If the current rendering task is a prerequisite for multiple subsequent rendering tasks, increase the priority of the rendering task.

6. The method according to claim 1, wherein: The method of determining the preference of digital human rendering and scene rendering in the video ringback tone of the call object according to the user attribute of the call object, and adjusting the digital human rendering parameters and the scene rendering parameters based on the preference of digital human rendering and scene rendering in the video ringback tone of the call object includes: Based on historical user experience data, use the decision tree algorithm to train the model, predict changes in user experience indicators, and adjust task priorities in advance before user experience indicators decrease; During the execution of rendering tasks, the execution status and user experience indicators of each rendering task are continuously monitored, and the rendering task allocation strategy is adjusted based on real-time data; Obtain the device performance parameters of the call object through the device performance detection library navigator.hardwareConcurrency, including the number of CPU cores, memory size, and GPU type; Based on the device performance parameters, set the performance baseline to determine whether the device can handle high, medium and low levels of rendering complexity; Use the network monitoring tool SpeedtestAPI to obtain the network bandwidth and connection speed data of both parties in the call. If the network bandwidth of either party is lower than the predetermined threshold, reduce the resolution and frame rate of the video call; Add user preference options in the app settings, including video resolution preferences and options to prioritize image quality or speed, and use user profiles stored locally or in the cloud to obtain user preferences; Based on user preferences and environmental sensor data, the brightness and background noise level of the call recipient's environment are determined, and the brightness and volume settings of the video ringback tone are adjusted; Based on the user attributes and historical usage data of the caller, the decision tree algorithm is used to train the model and predict the caller's video ringback tone style preference. User attributes include age, gender, and hobbies. Adjust the rendering style of digital people and scenes according to the video ringback tone style preference of the caller; Based on the device performance, network conditions, video ringtone style preferences and environmental data of the call partner, a recurrent neural network algorithm is used for model training to obtain the optimal digital human rendering parameter settings and scene rendering parameter settings for the call. The digital human rendering parameters include the digital human model detail level, digital human texture resolution, digital human material complexity and digital human dynamic effect requirements. The scene rendering parameters include scene size and scene environment effects.

7. The method according to claim 1, wherein: The method monitors the network bandwidth utilization and latency through Nagios, performs data pre-fetching and data synchronization based on the monitoring results, uses LRU and FIFO to manage the cache servers corresponding to each computing node, dynamically allocates rendering tasks based on the status of the cache servers, and continuously optimizes the network environment and cache server configuration based on the cache server load and performance indicators, including: The network monitoring tool Nagios is used to continuously monitor the bandwidth utilization and latency of the current network. If changes in the network environment are detected, the pre-fetch mechanism is triggered. Changes in the network environment include bandwidth utilization rising by a value greater than a preset rising value threshold or latency increasing by a value greater than a preset increase threshold. According to the pre-fetch trigger conditions, a data pre-fetch strategy is adopted to obtain historical access patterns and user behavior data related to the task, and a decision tree algorithm is used for model training to determine the resource data that should be pre-loaded into the cache; If it is determined that resource data needs to be pre-fetched, a timestamp mechanism is applied to synchronize the source data with the data in the cache; After completing data pre-fetching, the data in the cache is managed using the data expiration strategy in the cache strategy. The data expiration strategy includes the least used LRU and the first-in-first-out FIFO. Based on the node allocation score evaluation formula, rendering tasks are dynamically allocated according to the resource status and task requirements of the cache server; If it is detected that the resource load of the designated server is higher than the preset second load threshold, the rendering task to be allocated is reallocated to a server whose load is lower than the preset third load threshold; Evaluate the performance of the cache server based on the hit rate and service quality indicators in the cache server's performance indicators. If the cache server's performance does not meet the standards, expand the cache server's performance level by adjusting the cache size or increasing the number of servers. According to the proportion of time the cache server is operating normally, the cache server availability index is evaluated and the response time data is obtained, where the response time is the time from when the user initiates the request to when the cache server responds; Continuously evaluate the user experience quality during the rendering task execution process through response time and cache server availability indicators. If user experience problems are detected, optimize the network environment and cache server configuration; Apply a fault recovery strategy. When a cache server fails, switch to a backup server or reallocate resources. The fault recovery strategy includes a backup mechanism and a fault tolerance mechanism.

8. The method according to claim 1, wherein: The health status of each computing node of the rendering server is regularly detected. If the response time of the computing node exceeds a preset threshold, it is identified as an unhealthy node and task reallocation is triggered, including making decisions based on performance evaluation data and task classification data to migrate tasks of the unhealthy node to other healthy nodes with matching performance, including: Implement a scheduled task mechanism, set the rendering task detection interval, detect the health detection indicators of each computing node of the rendering server, and evaluate the health status of each computing node. The health detection indicators include CPU usage and memory usage; Regularly obtain the response time data of each node, and use the monitoring tool Prometheus combined with Grafana to determine whether the node is within the preset response time range; For nodes with abnormal response time or other abnormal health detection indicators, retry detection is performed. If the number of consecutive detection failures is greater than the preset threshold, the node is determined to be an unhealthy node; Obtain the error log of unhealthy nodes, use the abnormal information to assist in determining the fault type of the node, and determine the corresponding fault repair measures; Obtain task information related to unhealthy nodes from the task queue and the mapping information between tasks and computing nodes, and evaluate the current status of the task; For each task to be migrated, the performance matching degree between the migration task and each normally running node is calculated based on the resource requirements of the migration task and the resource supply capacity of the node. The resource requirements include CPU intensiveness and memory intensiveness. Based on the urgency of the task, resource requirements, and user needs, a decision tree algorithm is used to train the model and evaluate the priorities of all tasks to be migrated, including high priority, medium priority, and low priority. The migration strategy is defined based on the task priority and performance matching metric. The migration strategy includes determining the order of migration tasks and selecting the best target node. High-priority migration tasks are migrated first and migrated to the node with the highest performance matching degree. According to the migration strategy, the tasks of unhealthy nodes are migrated to other healthy nodes with matching performance.

9. The method according to claim 1, wherein: The audio and video call link is established between the rendering server and the user terminal through the Coturn server, and the rendered digital human and scene are transmitted by the audio and video call link after the rendering task is completed, including: According to the network environment of the rendering server and user terminal, configure the STUN service through the Coturn server to obtain the public IP and port information of the rendering server and user terminal to determine whether there are NAT or firewall restrictions; If both the user terminal and the rendering server are behind NAT, use the TURN service of the Coturn server to establish a relay to transmit data between devices behind different NATs; The security protocol DTLS is used to encrypt the audio and video data streams, and an audio and video transmission solution is obtained to protect the communication link; Use the WebRTC protocol to establish a real-time audio and video call link between the rendering server and the user terminal to evaluate the link bandwidth and delay, and optimize the quality of audio and video transmission based on the link bandwidth and delay conditions; Adjust audio and video properties based on the transmission quality of the real-time call link, including reducing resolution or frame rate, to obtain the best call quality that is suitable for the current network environment; The processing capacity of the rendering server is judged through the performance parameters of the rendering server, and the complexity of the rendering task is determined based on the processing capacity of the rendering server. The performance parameters of the rendering server include CPU performance and GPU performance; Obtain the device type and operating system information of the user terminal, evaluate the compatibility and display effect of the rendered content on the user terminal, and determine the optimal rendering parameters for the user device; Through the WebRTC session control protocol SessionDescriptionProtocol, a session management mechanism for audio and video calls is established to synchronously transmit the digital human model and scene rendering tasks on the call link; After the rendering task is completed, the rendered digital people and scenes are transmitted to the user terminal through the established secure and encrypted audio and video call link, realizing real-time audio and video interaction.

10. The method according to claim 1, wherein: The method constructs a node performance prediction model based on the historical performance data of each computing node of the rendering server to obtain the load and performance changes of each computing node of the rendering server in a preset time period in the future, determines the performance bottleneck of each computing node, and performs the redistribution of rendering tasks in advance, including: Obtain key performance indicator data of each computing node of the rendering server, including CPU usage, GPU usage, memory usage, disk I / O throughput, and network throughput data; Based on historical rendering task data, identify the computing resource requirements of different types of rendering tasks and the node resource consumption patterns. Different types of rendering tasks include CPU-intensive, GPU-intensive, and memory-intensive tasks. The acquired key performance indicator data is sorted into timestamp sequences, and the ARIMA algorithm is used for model training to predict the key performance indicator values ​​within a preset time period in the future. Based on the predicted trend line, seasonal component and residual, the periodic fluctuation and time dependence of node performance are determined; Based on the historical task execution data, the correlation between the task execution mode and the node performance is determined by calculating the Pearson correlation coefficient. The historical task execution data includes task duration, resource consumption, and queue waiting time. Obtain system event data from system maintenance logs, hardware replacement records, and abnormal event logs. For each obtained system event, associate the computing node performance indicator data within the corresponding time period. According to the performance indicator changes before and after the system event, identify the impact of the system event on the computing node performance and task execution. The system event data includes the date and time of maintenance, maintenance activity description, replacement parts and reasons for hardware replacement, and the time, type, and impact range of abnormal events. Obtain key hardware configuration information of each computing node of the rendering server, including the number of CPU cores, CPU model, GPU model and computing power, memory capacity, disk type and disk capacity, and network interface type; Correlate the acquired hardware configuration information with the performance indicator data of the node, use the variance analysis method to determine the correlation between the hardware configuration parameters and the node performance indicators, and identify the hardware configuration factors whose impact on the node performance is greater than the preset threshold; Based on hardware configuration, performance indicators, task characteristics, and system event log data, the random forest algorithm is used to train the model and build a node performance prediction model to predict the performance changes of computing nodes in a preset time period in the future. According to the prediction results of the node performance prediction model, potential performance bottleneck areas are identified, wherein the performance bottleneck is a node with degraded performance; Based on the prediction results of the node performance prediction model and the performance bottleneck prediction, the distribution of rendering tasks is adjusted in advance. If a node is predicted to face a performance bottleneck, the rendering tasks are reallocated in advance.