A method and device for scheduling online cloud rendering tasks and offline cloud rendering tasks

Abstract

The application discloses a scheduling method and device for online cloud rendering tasks and offline cloud rendering tasks. The method first acquires task information of rendering materials in offline rendering tasks to be processed, and processes the rendering materials based on a first prediction model to obtain a predicted rendering completion time length of the offline rendering tasks on at least one type of graphic processing unit (GPU). Meanwhile, based on historical resource consumption data of online cloud rendering tasks, the second prediction model is processed to obtain a predicted resource amount required by the online cloud rendering tasks in a future preset period. Then, based on the predicted resource amount and a total resource amount of a resource pool, an available resource amount available for scheduling the offline rendering tasks in the future preset period is determined. Finally, according to the available resource amount and the predicted rendering completion time length, a target offline rendering task is selected from the offline rendering tasks to be processed, and the target offline rendering task is scheduled to be executed in the future preset period.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and device for scheduling online cloud rendering tasks and offline cloud rendering tasks. Background Technology

[0002] With the rapid development of virtual reality and 3D digital technologies, the demand for cloud rendering services such as cloud gaming, cloud AR, and 3D animation rendering has exploded. Cloud rendering services are divided into online interactive services with stringent requirements for low latency and high concurrency, and offline batch processing services with core requirements for task execution integrity. How to efficiently allocate GPU computing resources and balance service quality assurance with resource utilization improvement has become a core issue that the cloud rendering industry urgently needs to solve.

[0003] Currently, the industry generally adopts a resource pool isolation scheme for GPU computing power scheduling in cloud rendering services. This involves dividing the GPU cluster into two isolated online and offline service resource pools, with each pool being independently scheduled and maintained. At the same time, a monitoring system collects node load data and dynamically adjusts the number of nodes in the two resource pools based on load changes, thereby optimizing cluster resource utilization.

[0004] The existing solutions described above have many insurmountable drawbacks. The initial resource pool can only be statically divided based on historical experience, which inherently leads to unreasonable resource allocation and makes it prone to resource imbalance, resulting in a ceiling on the cluster's computing power utilization. Furthermore, adjusting nodes between two isolated resource pools requires operations such as business migration and environment redeployment, which are not only time-consuming and costly but also pose a risk of business interruption, failing to achieve flexible and rapid allocation of computing power. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a method and device for scheduling online cloud rendering tasks and offline cloud rendering tasks.

[0006] The embodiments of this application disclose the following technical solutions: In a first aspect, embodiments of this application provide a method for scheduling online cloud rendering tasks and offline cloud rendering tasks, the method comprising: Obtain task information for offline rendering tasks to be processed, including rendering materials; The rendering material is processed by the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processing unit (GPU). Based on historical resource consumption data of online cloud rendering tasks, the predicted resource amount required by the online cloud rendering tasks in a future preset period is predicted by a second prediction model. Based on the predicted resource quantity and the total resource quantity of the resource pool, the available resource quantity that can be used to schedule offline rendering tasks within the future preset period is determined. Based on the available resources and the predicted rendering completion time, select a target offline rendering task from the offline rendering tasks to be processed. The target offline rendering task is scheduled to be executed within the preset future period.

[0007] In one possible implementation, processing the rendered material using a first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processing unit (GPU) includes: Analyze the rendered material to obtain at least one feature parameter of the rendered material; By inputting the at least one feature parameter and the benchmark information of the computing power of the corresponding type of GPU in the GPU computing power information database into the first prediction model, the predicted rendering completion time of the offline rendering task on the target type of GPU is obtained.

[0008] In one possible implementation, the step of predicting the predicted resource amount required by the online cloud rendering task within a future preset period based on historical resource consumption data of the online cloud rendering task using a second prediction model includes: Input the historical resource consumption data into the second prediction model to obtain the demand trend of GPU resources for online cloud rendering tasks in one or more future task cycles; Based on the demand trend, the number of GPUs required for online cloud rendering tasks within the future preset period is determined as the predicted resource quantity.

[0009] In one possible implementation, selecting a target offline rendering task from the pending offline rendering tasks based on the available resource quantity and the predicted rendering completion time includes: The offline rendering tasks to be processed are sorted according to the rendering completion time of the offline rendering tasks on at least one type of GPU. Based on the sorting results and the available resources, one or more offline rendering tasks whose rendering completion time meets the resource constraints are selected in sequence as the target offline rendering tasks.

[0010] In one possible implementation, after scheduling the target offline rendering task to execute within the future preset period, the method further includes: Monitor the load data of the GPU nodes executing the target offline rendering task; After the target offline rendering task is completed, the corresponding GPU node resources are released based on the load data.

[0011] In one possible implementation, the method further includes: Construct a GPU computing power information database; the GPU computing power information database includes computing power performance benchmark information corresponding to various types of GPUs supported by the resource pool.

[0012] In one possible implementation, the first prediction model is a task duration prediction model trained based on the Support Vector Regression (SVR) algorithm, and the second prediction model is an online resource prediction model trained based on the Prophet time series prediction algorithm.

[0013] Secondly, embodiments of this application disclose a scheduling device for online cloud rendering tasks and offline cloud rendering tasks, the device comprising: The acquisition module is used to acquire task information of the offline rendering task to be processed, the task information including rendering materials; The first prediction module is used to process the rendering material through the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processor GPU. The second prediction module is used to predict the amount of resources required by the online cloud rendering task within a future preset period based on the historical resource consumption data of the online cloud rendering task and through the second prediction model. The determination module is used to determine the amount of available resources that can be used to schedule offline rendering tasks within the future preset period based on the predicted resource amount and the total resource amount of the resource pool. The filtering module is used to select a target offline rendering task from the offline rendering tasks to be processed based on the available resource quantity and the predicted rendering completion time. The scheduling module is used to schedule the target offline rendering task to be executed within the preset future period.

[0014] In one possible implementation, the first prediction module is specifically used to parse the rendering material to obtain at least one feature parameter of the rendering material; input the at least one feature parameter and the computing power performance benchmark information of the corresponding type of GPU in the GPU computing power information database into the first prediction model to obtain the predicted rendering completion time of the offline rendering task on the target type GPU.

[0015] In one possible implementation, the second prediction module is specifically used to input the historical resource consumption data into the second prediction model to obtain the demand trend of GPU resources for online cloud rendering tasks in one or more future task cycles; and to determine the number of GPUs required for online cloud rendering tasks in the future preset cycle as the predicted resource quantity based on the demand trend.

[0016] In one possible implementation, the filtering module is specifically used to sort the offline rendering tasks to be processed according to their rendering completion time on at least one type of GPU; and, based on the sorting result and the available resources, sequentially select one or more offline rendering tasks whose rendering completion time meets the resource constraints as the target offline rendering tasks.

[0017] In one possible implementation, after the target offline rendering task is scheduled to be executed within the preset future period, the scheduling module is further configured to monitor the load data of the GPU node executing the target offline rendering task; after the target offline rendering task is completed, the corresponding GPU node resources are released according to the load data.

[0018] In one possible implementation, the device further includes a construction module for constructing a GPU computing power information database; the GPU computing power information database includes computing power performance benchmark information corresponding to various types of GPUs supported by the resource pool.

[0019] In one possible implementation, the first prediction model is a task duration prediction model trained based on the Support Vector Regression (SVR) algorithm, and the second prediction model is an online resource prediction model trained based on the Prophet time series prediction algorithm.

[0020] Thirdly, embodiments of this application disclose a control device, including a processor and a memory, wherein the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to complete the scheduling method for online cloud rendering tasks and offline cloud rendering tasks as described in any of the first aspects.

[0021] Fourthly, embodiments of this application disclose a computer-readable storage medium, characterized in that it stores a computer program, which is loaded by a processor to execute the scheduling method for online cloud rendering tasks and offline cloud rendering tasks as described in any of the first aspects.

[0022] This application provides a method and apparatus for scheduling online and offline cloud rendering tasks. The method first acquires task information of rendering materials in the offline rendering task to be processed, and then processes the rendering materials using a first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of GPU. Simultaneously, based on historical resource consumption data of the online cloud rendering task, a second prediction model is used to obtain the predicted resource amount required by the online cloud rendering task within a future preset period. Then, based on the predicted resource amount and the total resource amount of the resource pool, the available resource amount that can be used to schedule the offline rendering task within the future preset period is determined. Finally, based on the available resource amount and the predicted rendering completion time, a target offline rendering task is selected from the offline rendering tasks to be processed, and the target offline rendering task is scheduled to execute within the future preset period.

[0023] This application's embodiments, by predicting the resource demands of online cloud rendering tasks within a preset future period, transform the existing technology's passive resource adjustment based on real-time load into proactive resource planning. This effectively solves the resource scheduling lag problem caused by the tidal fluctuations in online service traffic, allowing for the pre-reservation of necessary computing resources for online services and ensuring service quality. Simultaneously, by predicting the rendering completion time of offline rendering tasks and combining this with the available schedulable resources to select suitable target offline tasks, it ensures that the scheduling and execution of offline tasks do not encroach on the resources required by online services. It also maximizes the reuse of idle computing power within the resource pool, effectively solving the problems of unreasonable resource allocation and low computing power utilization in existing technologies. While ensuring the service quality of online services, it significantly improves the overall utilization efficiency of GPU computing resources. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a scheduling method for online cloud rendering tasks and offline cloud rendering tasks provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for scheduling online and offline cloud rendering tasks provided in this application embodiment; Figure 3 This is a schematic diagram of a scheduling device for online cloud rendering tasks and offline cloud rendering tasks provided in an embodiment of this application. Detailed Implementation

[0026] As described earlier, existing GPU computing power scheduling for cloud rendering services generally adopts a technical solution of isolating online and offline service resource pools for independent scheduling and operation. This solution relies solely on monitoring systems to collect node load data and dynamically adjust the number of nodes in the two resource pools based on continuous periodic load changes to optimize computing resource allocation and utilization. However, this approach not only relies on statically partitioning the initial resource pools based on historical experience, inherently leading to unreasonable resource allocation and a ceiling that is difficult to overcome in terms of overall cluster GPU computing power utilization, but also requires business migration and runtime environment reconfiguration when adjusting nodes across resource pools. The complex deployment and other operations have drawbacks such as long adjustment cycles, high computing power and operation and maintenance costs, and high risk of business interruption. It cannot achieve flexible and rapid allocation of computing resources. At the same time, for the significant tidal traffic fluctuations of online cloud rendering services, the passive scaling mechanism based on real-time load has obvious response lag. It cannot replenish computing power in time during traffic peaks to ensure the service quality of online services, nor can it release computing power in time during traffic troughs to avoid resource waste. In addition, the isolated deployment architecture cannot reuse the fragmented idle computing power generated during the operation of online services, further exacerbating the waste of GPU computing power resources.

[0027] To address this technical problem, embodiments of this application provide a method and apparatus for scheduling online and offline cloud rendering tasks. The method first acquires task information of the rendering materials in the offline rendering task to be processed, and then processes the rendering materials using a first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of GPU. Simultaneously, based on historical resource consumption data of the online cloud rendering task, a second prediction model is used to obtain the predicted resource amount required by the online cloud rendering task within a future preset period. Then, based on the predicted resource amount and the total resource amount of the resource pool, the available resource amount that can be used to schedule the offline rendering task within the future preset period is determined. Finally, according to the available resource amount and the predicted rendering completion time, a target offline rendering task is selected from the offline rendering tasks to be processed, and the target offline rendering task is scheduled to execute within the future preset period.

[0028] This application's embodiments, by predicting the resource demands of online cloud rendering tasks within a preset future period, transform the existing technology's passive resource adjustment based on real-time load into proactive resource planning. This effectively solves the resource scheduling lag problem caused by the tidal fluctuations in online service traffic, allowing for the pre-reservation of necessary computing resources for online services and ensuring service quality. Simultaneously, by predicting the rendering completion time of offline rendering tasks and combining this with the available schedulable resources to select suitable target offline tasks, it ensures that the scheduling and execution of offline tasks do not encroach on the resources required by online services. It also maximizes the reuse of idle computing power within the resource pool, effectively solving the problems of unreasonable resource allocation and low computing power utilization in existing technologies. While ensuring the service quality of online services, it significantly improves the overall utilization efficiency of GPU computing resources.

[0029] The technical solutions provided in this application are widely applicable to all cloud rendering scenarios equipped with various GPU computing hardware. They are primarily used by cloud rendering operation service providers, film and animation production companies, and digital twin platform operators that have unified GPU computing infrastructure. The underlying hardware carriers include professional-grade rendering GPU clusters in local data centers, lightweight GPU devices for edge computing nodes, elastically scalable GPU resource pools in the cloud, as well as supporting virtualization hardware platforms, distributed storage arrays, and dedicated computing power scheduling servers.

[0030] For applications that simultaneously support online rendering services such as cloud gaming and real-time cloud AR / VR interaction, as well as offline rendering services such as 3D animation rendering, film and television special effects production, and batch baking of 3D scenes, this solution can directly achieve hybrid deployment and collaborative scheduling of the two types of services on a unified GPU hardware cluster. There is no need to build separate physically isolated hardware resource pools for the two types of services. It can reserve sufficient GPU computing power for low-latency online services in advance through online resource prediction, ensuring the stability of hardware resource supply. At the same time, it can accurately match the rendering time requirements of offline tasks with the hardware idle computing power window, maximizing the reuse of fragmented hardware computing power generated by the tidal fluctuations of online services, and significantly improving the overall utilization efficiency of GPU hardware devices.

[0031] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0032] See Figure 1 , Figure 1This is a flowchart illustrating a scheduling method for online and offline cloud rendering tasks provided in an embodiment of this application. The executing entity of this method can be a server, desktop computer, or other device capable of computational functions. The following description uses a computing device as the executing entity, and the method includes: S101: The computing device obtains task information for the offline rendering task to be processed.

[0033] In this embodiment, the computing device is the core hardware carrier that carries the scheduling method. Specifically, it can be the central scheduling server of the cloud rendering platform, the management and control node of the computing power cluster, the edge computing device that integrates rendering scheduling function, or the computing power management unit of the cloud elastic computing power platform, which has the core capabilities of data reception, parsing, storage and preprocessing.

[0034] The computing device receives offline rendering tasks submitted by users through a pre-defined standardized task submission interface, and simultaneously obtains all task information corresponding to the task. Offline rendering tasks are non-real-time batch rendering tasks, distinct from real-time interactive online cloud rendering tasks. Examples include 3D animation frame rendering, film and television special effects rendering, 3D digital twin scene baking, and batch rendering of architectural renderings. Offline rendering tasks do not have stringent requirements for single-interaction latency; their core requirement is to ensure the integrity of rendering execution and the accuracy of the output results.

[0035] Users can submit offline rendering tasks through various channels such as the cloud rendering platform's web client, desktop client, and standardized API interface. After receiving the task submission request, the computing device will retrieve, verify, and store the corresponding task information.

[0036] The task information acquired by the computing device includes rendering assets, and may also include user-specified rendering configuration auxiliary information. Rendering assets are the core set of source files supporting offline rendering tasks, serving as the foundational data for the 3D rendering engine's rendering calculations. These include source data directly related to the rendering scene, such as 3D scene project files, model mesh files, material texture files, lighting configuration files, animation keyframe files, and scene proxy files. The scene complexity, data volume, and parameter configuration of the rendering assets directly determine the computing power requirements, computational complexity, and final execution time of the rendering task.

[0037] In addition to the core rendering materials, task information can also include user-defined rendering parameters, such as rendering engine type, rendering resolution, sampling rate, rendering frame range, output file format, specified GPU specification type, expected task completion time or task priority, etc. User-defined rendering parameters are synchronously acquired and stored by the computing device, providing complete basic data support for subsequent task duration prediction, task selection and scheduling execution.

[0038] S102: The computing device processes the rendering material through the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processor GPU.

[0039] The first prediction model is an offline rendering task duration prediction model pre-trained and optimized based on the Support Vector Regression (SVR) algorithm. It uses the characteristics of rendering materials, the type of GPU used, and the actual rendering completion time of historical offline rendering tasks as labeled training data. By constructing a regression hyperplane, it achieves accurate prediction of the rendering duration of unknown tasks. The predicted rendering completion time refers to the estimated total time taken for an offline rendering task from starting the rendering process and loading rendering materials to completing the rendering calculations for all target frames and outputting the final result file. It is used to determine whether the task can be fully executed within a preset future period.

[0040] In this embodiment, the computing device performs structured analysis and feature extraction on the rendering materials, adapts to the project file formats of mainstream rendering engines such as Blender, Maya, or 3ds Max, and automatically identifies and extracts core task feature data that can characterize the computational complexity of the rendering task, including but not limited to the total number of triangles in the 3D scene, the number and resolution of material textures, the number and type of light sources, the total number of animation keyframes, the range of rendering frames, and the sampling rate, etc., to complete the standardized preprocessing of the rendering materials and provide standardized basic data for the input of the first prediction model.

[0041] After feature extraction of the rendered materials is completed, the computing device retrieves a pre-built GPU computing power information database to obtain benchmark computing performance information for at least one type of GPU supported by the resource pool. The GPU computing power information database is a structured database that centrally stores the rendering performance parameters of all schedulable GPU devices within the resource pool. The benchmark computing performance information stored in the database includes core parameters that quantify the GPU's rendering computing capabilities, such as the corresponding GPU's single-precision floating-point arithmetic capability, number of rendering cores, memory bandwidth, benchmark rendering efficiency for similar rendering scenes, and historical average computing power consumption data. The computing device inputs the pre-processed task feature data and the computing performance benchmark information corresponding to the target type GPU into the pre-trained first prediction model. Through the regression calculation processing of the model, the predicted rendering completion time of the offline rendering task on the corresponding type of GPU is directly output. If there are multiple schedulable types of GPUs in the resource pool, or if the user specifies multiple alternative GPU specifications, the computing device can input the computing performance benchmark information of different types of GPUs in batches to obtain the predicted rendering completion time of the offline rendering task on multiple types of GPUs at once. This enables batch duration prediction in multiple hardware scenarios and adapts to the scheduling requirements of heterogeneous GPU hardware in the resource pool.

[0042] This application embodiment achieves pre-precise prediction of offline rendering task execution time through a pre-trained first prediction model, solving the problems of uncontrollable offline rendering task execution time and inability to plan and schedule in advance in existing technologies. Existing technologies can only obtain the actual time consumption after the task is completed, and cannot quantitatively evaluate the task execution cycle before scheduling, which easily leads to problems such as task execution timeouts and encroachment on the computing resources reserved for online services. However, this application embodiment can complete the accurate prediction of the duration before task scheduling, providing a reliable quantitative basis for subsequent task selection, ensuring that the selected target offline rendering task can be fully executed within the future preset cycle without affecting the resource supply of online services.

[0043] To further improve the prediction accuracy of offline rendering task completion time and ensure the accuracy and reliability of subsequent task scheduling decisions, this embodiment of the invention also pre-constructs a standardized GPU computing power information database and optimizes the execution process of offline rendering task duration prediction based on this GPU computing power information database.

[0044] The GPU computing power information database is a structured database that centrally stores and manages rendering computing power information related to all schedulable GPU hardware within the resource pool. The database includes benchmark computing power information for each type of GPU supported by the resource pool. The benchmark information stored in the database is a standardized set of parameters that quantifies the 3D rendering computing capabilities of different types of GPUs. Specifically, it can include three main categories: hardware core specifications of each type of GPU, rendering performance calibration data, and historical task test data. During the construction process, the computing device first performs a full inventory of all deployed and schedulable GPU devices within the resource pool, identifying the model, architecture type, and hardware specifications of each GPU, and collecting core hardware parameters such as single-precision floating-point arithmetic capability, number of rendering cores, memory capacity, and memory bandwidth. Subsequently, for each type of GPU, the computing device used a standardized 3D rendering test scenario to calibrate the rendering performance, and obtained calibration data such as the baseline rendering frame rate, the number of triangles that can be processed per unit time, and the average rendering time per frame under different complexity scenarios. At the same time, it pulled the actual running data of this type of GPU in historical offline rendering tasks, including the actual rendering efficiency under different scene characteristics, the pattern of computing power resource consumption, and the statistical data of task completion time.

[0045] After normalizing, filtering outliers, and validating the three types of data collected, the computing device categorizes and stores the data in a structured manner according to GPU type, completing the initial construction of the GPU computing power information database. During the operation of the resource pool, the computing device will also update and maintain the computing power performance benchmark information in the GPU computing power information database in real time based on the deployment of new GPU devices, the retirement of hardware devices, and updates to historical task data, ensuring that the data in the database is completely matched with the actual hardware environment of the resource pool and the actual rendering performance of the GPUs.

[0046] After completing the construction of the GPU computing power information database, the computing device parses the rendering material to obtain at least one feature parameter of the rendering material; and inputs at least one feature parameter and the computing power performance benchmark information of the corresponding type of GPU in the GPU computing power information database into the first prediction model to obtain the predicted rendering completion time of the offline rendering task on the target type of GPU.

[0047] Specifically, the computing device performs standardized parsing and feature extraction processing on the rendering materials acquired from the offline rendering task to obtain at least one feature parameter for duration prediction.

[0048] Feature parameters are core quantitative indicators that directly characterize the computational complexity and computing power consumption of offline rendering tasks. They are the core input items for the first prediction model to achieve accurate prediction of rendering time. Specifically, they may include the total number of triangles in the 3D scene corresponding to the rendering material, the number and resolution of material textures, the number and type of light sources, the total number of animation keyframes, the target rendering resolution, the sampling rate, and the rendering frame range. In the specific implementation process, the computing device first identifies the rendering engine type corresponding to the rendering material, adapts to the project file formats of mainstream 3D rendering engines such as Blender, Maya, 3ds Max, and Cinema4D, and calls the corresponding parsing plugin to complete the full parsing of the rendering material. Subsequently, the above-mentioned core feature parameters are automatically extracted from the parsed project file, and the extracted feature parameters are standardized preprocessed and normalized to eliminate data differences caused by different rendering engines and different parameter formats, generating a standardized feature parameter set that conforms to the input specifications of the first prediction model. If the rendering material includes segmented rendering requirements for multiple shots and multiple scenes, the computing device can also extract the corresponding feature parameters for each rendering segment to achieve segmented rendering time prediction, further improving the granularity of the prediction results.

[0049] After extracting the feature parameters of the rendering materials, the computing device inputs the standardized feature parameters, along with the benchmark computing performance information of the corresponding type of GPU in the GPU computing power information database, into the pre-trained first prediction model. Through regression calculation processing of the model, the predicted rendering completion time of the offline rendering task on the target type of GPU is output. In the specific execution process, if the user specifies the target type of GPU, the computing device directly retrieves the benchmark computing performance information corresponding to that type of GPU from the GPU computing power information database and inputs it into the model along with the feature parameters to complete the time prediction for a single type of GPU. If there are multiple schedulable heterogeneous types of GPUs in the resource pool, the computing device can batch retrieve the benchmark computing performance information of all available types of GPUs in the GPU computing power information database, combine them with the feature parameters respectively, and input them into the first prediction model to obtain the predicted rendering completion time of the offline rendering task on multiple different types of GPUs at once, providing a multi-dimensional decision basis for subsequent task scheduling. This embodiment provides a standardized hardware performance benchmark for the first prediction model through a pre-built GPU computing power information library. Combined with the complexity characteristics of the rendering materials themselves, it provides complete input support for duration prediction from two core dimensions: task computing requirements and hardware computing capabilities. This effectively eliminates prediction bias caused by differences in the performance of different GPU hardware, controlling the prediction error of rendering completion time to within 10%. This not only ensures the accuracy of the prediction results but also provides a reliable quantitative basis for subsequent calculation of available resources, selection of target tasks, and scheduling execution. It ensures that the scheduling of offline rendering tasks can accurately match the idle computing power window of the resource pool, maximizing the utilization rate of GPU computing power resources without affecting the quality of online business services.

[0050] S103: The computing device uses a second prediction model to predict the amount of resources required for online cloud rendering tasks within a future preset period, based on historical resource consumption data of online cloud rendering tasks.

[0051] Historical resource consumption data refers to the quantifiable computing power resource usage time-series data generated during the full operation of online cloud rendering services within a preset historical statistical period. This data can include continuous time-series data such as GPU core utilization, video memory usage, number of online concurrent users, average computing power consumption per user, and peak / valley time distribution, statistically analyzed at a preset time granularity. This data forms the foundation for resource demand prediction. The second prediction model is an online resource demand prediction model pre-trained and optimized based on the Prophet time-series prediction algorithm. It possesses core capabilities such as time-series data trend mining, period fitting, and outlier compatibility, and serves as the core computational vehicle for resource demand prediction in this step.

[0052] The future preset period is the smallest time unit for pre-configured resource prediction and task scheduling, which can be flexibly configured to 1 hour, 4 hours, 1 calendar day, etc., depending on the business scenario. The predicted resource quantity refers to the minimum total amount of GPU computing power resources reserved to ensure the stable operation of online cloud rendering services within the future preset period, calculated by the second prediction model, and is characterized by quantifiable computing power units such as the number of GPU cards, the percentage of GPU cores, and the capacity of video memory.

[0053] The computing device pulls the full historical resource consumption data of online cloud rendering services within a preset historical statistical period from the monitoring system of the resource pool. It performs preprocessing operations such as automatic missing value completion, extreme outlier filtering, and time granularity alignment of time series data on the pulled raw data to generate a standardized time series dataset that conforms to the input specifications of the second prediction model, providing reliable basic data for subsequent prediction calculations.

[0054] After preprocessing historical resource consumption data, the computing device inputs the standardized time-series dataset into the pre-trained second prediction model. Through in-depth data mining and fitting calculations of the time-series data, the second prediction model outputs the predicted resource volume for online cloud rendering services within a preset future period. During the specific model calculation process, the second prediction model automatically extracts three core features from the input time-series data: first, long-term trend characteristics, i.e., the overall resource demand change pattern of online cloud rendering services over time, such as the overall upward trend in computing power demand due to user growth; second, periodic fluctuation characteristics, i.e., the fixed periodic change pattern of online service resource consumption, including daily peak-to-valley fluctuations, weekday-to-weekday fluctuations, and monthly / quarterly periodic changes, accurately matching the tidal traffic characteristics of online services; and third, special event fluctuation characteristics, i.e., the sudden changes in resource demand brought about by special dates such as holidays, large-scale online events, and business version updates.

[0055] The second prediction model, based on three extracted core features and combined with the time attributes corresponding to the future preset period, completes multi-dimensional fitting calculations and ultimately outputs the predicted resource quantity within the future preset period, including the average resource demand and peak resource demand within the period. In this embodiment, the peak resource demand is preferentially used as the benchmark value for subsequent available resource quantity calculation to maximize the supply of computing power and service stability during peak online business periods. If the business scenario requires long-term resource planning, the computing device can adjust the output parameters of the second prediction model to output the predicted resource quantity for multiple consecutive preset periods at once, realizing batch, long-term scheduling planning for offline rendering tasks.

[0056] This application embodiment, through a pre-trained second prediction model, achieves proactive and accurate prediction of online cloud rendering service resource requirements within a preset future period. This solves the response lag problem inherent in existing passive resource adjustment mechanisms based on real-time load, transforming the existing "post-event resource allocation" into "pre-event resource planning," providing core support for ensuring the service quality of online services. In existing technologies, resource adjustments are only triggered after the service load increases to address the tidal fluctuations in online service traffic. This easily leads to service quality issues such as online service lag and excessive latency due to insufficient computing power supply during peak periods. However, this application embodiment can accurately predict resource requirements a full scheduling cycle in advance, reserving sufficient computing power resources for online services in advance. This fundamentally avoids the risk of service quality degradation caused by delayed computing power supply, and also avoids the waste of computing power caused by excessive resource reservation during off-peak periods.

[0057] To further improve the accuracy of online cloud rendering service computing resource demand forecasting and the feasibility of scheduling, and to clarify the quantitative execution standards for resource reservation, this application embodiment further refines and limits the calculation process of online cloud rendering task forecasting resource quantity. While absolutely guaranteeing the low latency service quality of online services, it also maximizes the release of reusable idle computing resources.

[0058] In one possible implementation, based on historical resource consumption data of online cloud rendering tasks, a second prediction model is used to predict the amount of resources required for online cloud rendering tasks within a future preset period, including: The computing device inputs historical resource consumption data into the second prediction model to obtain the demand trend of GPU resources for online cloud rendering tasks within one or more future task cycles. Based on the demand trend, the number of GPUs required for online cloud rendering tasks within a preset future cycle is determined as the predicted resource quantity.

[0059] Historical resource consumption data refers to the time-series data of full GPU resource usage for online cloud rendering services collected by the resource pool monitoring system at a preset fixed time granularity within a preset historical statistical period. The time granularity can be configured to 5 minutes, 15 minutes, or 1 hour according to the business scheduling accuracy requirements, and the historical statistical period can be configured to 3 months, 6 months, or 12 months according to the accumulation of business data. The data content specifically includes continuous and traceable time-series data such as the number of online concurrent users at each time node, the average GPU core utilization rate per user, the average video memory consumption per user, the overall core utilization rate and video memory utilization rate of the GPU cluster, and the extreme values ​​of resource consumption during business peak and valley periods.

[0060] The computing device pulls all raw historical resource consumption data from the monitoring system and performs standardized preprocessing on the data. For missing values ​​caused by data acquisition interruptions, historical averages for the same period are used to fill in the gaps. Extreme outliers caused by non-normal business scenarios such as business failures and temporary stress tests are filtered out. Simultaneously, the computing device performs unified time-granularity alignment on all data to eliminate data acquisition bias, ultimately generating a standardized time-series dataset that conforms to the input specifications of the second prediction model.

[0061] After data preprocessing, the computing device inputs the standardized time-series dataset into the pre-trained second prediction model. Through deep fitting and feature mining of the time-series data, the model outputs the demand trend of GPU resources for online cloud rendering tasks within one or more future task cycles. Specifically, the demand trend includes the long-term trend of GPU resource consumption in online cloud rendering, the daily and weekly cyclical fluctuation trends, and the sudden fluctuation trends on holidays and special operational days. It also covers the peak, average, and trough data of resource demand for each future task cycle, comprehensively depicting the full variation pattern of resource demand corresponding to the tidal traffic fluctuations of online services.

[0062] After obtaining the GPU resource demand trend, the computing device, based on this trend and combined with the standardized computing power benchmark of a single GPU device within the resource pool, quantitatively calculates the number of GPUs required for online cloud rendering tasks within a preset period, and determines this number of GPUs as the predicted resource quantity. In the specific calculation process, the computing device first retrieves the single-card computing power performance benchmark information corresponding to the GPU types used in the resource pool from a pre-built GPU computing power information database, determining the online service computing power load threshold that a single GPU device can stably bear, thus avoiding excessive online service rendering latency caused by GPUs running at full load.

[0063] Subsequently, the computing device uses the peak resource demand within a preset future period as the core calculation benchmark, combined with the stable load threshold of a single GPU, to calculate the minimum number of GPUs required to ensure stable operation during peak online service periods. Simultaneously, based on the service level agreement requirements of the online service, a preset percentage of redundant protection GPUs can be added to ultimately determine the predicted resource quantity within the preset future period. For example, if the demand trend shows that the peak resource demand for online services within the preset future period is 160 available GPU rendering cores, and a single GPU device can stably provide 80 available rendering cores, with a 10% redundancy protection ratio configured, the final calculated base number of GPUs is 2. Adding 0.2 redundant GPUs and rounding up, the predicted resource quantity is determined to be 3 GPUs. If the second prediction model outputs demand trends for multiple consecutive task periods, the computing device can calculate the predicted resource quantity for each corresponding period according to the time boundary of each task period, providing a complete quantitative basis for the batch planning and scheduling of long-cycle offline rendering tasks.

[0064] This application's embodiments refine the calculation process for predicted resource quantities, transforming the prediction of resource requirements for online services from a vague computing power ratio indicator to a quantifiable indicator of the number of GPUs that can be directly used for scheduling and execution. This solves the problems of vague resource reservation standards and poor schedulability in existing technologies, providing accurate and directly applicable decision-making basis for subsequent calculation of available resource quantities and screening and scheduling of offline tasks. Furthermore, this application's embodiments support simultaneous calculation of demand trend prediction and GPU quantity for multiple future task cycles, adapting to the scheduling needs of long-cycle offline rendering tasks and further improving the scenario adaptability and flexibility of the scheduling scheme.

[0065] S104: The computing device determines the amount of available resources that can be used to schedule offline rendering tasks within a future preset period based on the predicted resource amount and the total resource amount of the resource pool.

[0066] The total resource pool refers to the total amount of GPU computing power resources in the resource pool that can be normally scheduled, without hardware failures, and without pre-occupancy maintenance plans within a preset future period. It is usually represented by quantifiable computing power units such as the number of standardized physical GPUs, the total number of available GPU cores, and the total available video memory capacity. Its value can be obtained and updated synchronously in real time from the hardware management system of the resource pool. Available resources refer to the maximum amount of idle computing power resources in the resource pool that can be safely allocated to offline rendering tasks within a preset future period, provided that the computing power supply for online cloud rendering services is absolutely guaranteed.

[0067] In this embodiment, the computing device retrieves the total resource volume for a future preset period from the resource pool hardware management system and retrieves the predicted resource volume for the same period, verifying the validity of both data. After data alignment and verification, the basic calculation logic for the available resource volume is as follows: Available resource volume for scheduling offline rendering tasks within the future preset period = Total resource volume of the resource pool - Predicted resource volume for online cloud rendering services. In the detailed implementation, the computing device can also combine the operation and maintenance requirements of the resource pool with the service level agreement of the online services, and add preset security redundancy adjustment items to further strengthen the service guarantee defense line for online services. Specifically, the computing device can deduct a preset proportion of online service redundancy protection resources on top of the basic calculation logic. As an example, this redundancy can be configured as a fixed proportion of 5%-10% of the total resource volume, or a fixed value of GPU computing power resources, to cope with sudden traffic fluctuations in online services that exceed the predicted value. If the resource pool includes heterogeneous hardware environments with GPUs of different computing power specifications, the computing device will convert the GPUs of different specifications into standard computing power units based on a pre-built GPU computing power information database before performing calculations, ensuring the accuracy of computing power calculation in heterogeneous environments.

[0068] After the initial calculation is completed, the computing device will also perform boundary checks and dynamic updates on the results. If the total resource quantity is less than or equal to the predicted resource quantity in an extreme case, the available resource quantity will be set to 0, the scheduling of all offline tasks within this period will be suspended, and online services will be given priority to be fully guaranteed. If the predicted resource quantity for online services or the total resource quantity in the resource pool changes due to device status or model recalculation, the available resource quantity will be recalculated immediately, and the computing power boundary for subsequent task selection will be updated synchronously to ensure that the scheduling decision always matches the actual state of the resource pool.

[0069] This application's embodiments, through quantitative calculations, define clear and implementable computing power security boundaries for the hybrid deployment of online and offline rendering tasks, solving the problem in existing technologies where offline tasks easily encroach on online service computing power resources and service quality cannot be guaranteed in hybrid scheduling scenarios. Furthermore, this application's embodiments, based on accurate prediction of online service resource demands and combined with the total computing power of the resource pool, achieve pre-emptive and precise calculation of available idle computing power. By using the logic of "first reserving online resources, then calculating available offline resources," it avoids the risk of offline tasks encroaching on reserved online service resources, improving the low-latency service quality of online services, while maximizing the release of idle computing power within the resource pool, avoiding unnecessary resource reservation waste, and providing accurate and reliable quantitative basis for the subsequent selection and scheduling of offline rendering tasks.

[0070] S105: The computing device selects a target offline rendering task from the pending offline rendering tasks based on the amount of available resources and the predicted rendering completion time.

[0071] The target offline rendering task refers to the offline rendering task that, after being screened, meets all constraints and can be safely scheduled and executed within a future preset period. It is the object of subsequent scheduling and execution. When screening target offline rendering tasks, the constraints of not exceeding the upper limit of available resources and not exceeding the time boundary of the future preset period are taken into account to ensure that the screened tasks will not encroach on the computing resources reserved for online services and can be fully executed within the preset period.

[0072] Specifically, the computing device can sort all pending offline rendering tasks in descending order of the minimum GPU computing power required for execution, forming a queue of tasks to be screened. This reverse sorting prioritizes offline tasks with high computing power requirements, avoiding resource fragmentation issues caused by large computing power tasks remaining unschedulable for extended periods and small tasks filling the resource pool. After sorting, the available resources of the computing device serve as the computing power limit, and the duration of a future preset period serves as the time limit. Each offline rendering task is then subjected to constraint checks according to the sorted queue order. For example, the first check is a computing power constraint: the minimum GPU computing power required for the task's execution does not exceed the currently remaining available resources. The second check is a time constraint: the predicted rendering completion time for the task does not exceed the total duration of the future preset period. If a task passes the constraint checks, it is added to the target offline rendering task set. Simultaneously, the computing power required for the task is deducted from the available resources, the remaining available resources are updated, and the next task in the queue is checked. If a task fails the constraint validation, it is skipped, and subsequent tasks are validated until the queue is traversed or the remaining available resources are insufficient to support any pending tasks.

[0073] In practical applications, computing devices can also overlay custom filtering rules, such as prioritizing user-specified high-priority tasks, prioritizing tasks with earlier submission times, and prioritizing tasks with a higher match between predicted rendering completion time and preset cycles, to further adapt to the scheduling needs of different business scenarios; if the available resources are 0, the filtering process is terminated directly, and no target offline rendering tasks are selected, strictly adhering to the bottom line of ensuring online business resources.

[0074] To further improve the matching efficiency between offline rendering tasks and available computing power, avoid task execution timeouts and computing power resource fragmentation, and ensure that the target task can be executed safely and completely within a preset period, this embodiment of the invention also refines and limits the screening process for target offline rendering tasks.

[0075] In one possible implementation, selecting a target offline rendering task from the pending offline rendering tasks based on the available resource quantity and the predicted rendering completion time includes: The computing device sorts the offline rendering tasks to be processed based on their rendering completion time on at least one type of GPU. Then, based on the sorting results and available resources, the computing device sequentially selects one or more offline rendering tasks whose rendering completion time meets resource constraints as target offline rendering tasks.

[0076] The computing device uses the predicted rendering completion time of each offline rendering task as the core benchmark to standardize the sorting of the task queue. If a task specifies a target GPU type, the corresponding predicted time for that GPU is directly used as the sorting benchmark. If it is a heterogeneous GPU cluster and the task does not specify a GPU type, the shortest predicted time of the task on available GPUs or the predicted time of the mainstream GPUs in the resource pool is used as the unified benchmark. The sorting rules can be flexibly configured according to business needs. They can be sorted in ascending order of predicted time to prioritize scheduling short-cycle tasks and improve task completion throughput, or in descending order of predicted time to prioritize scheduling long-cycle tasks and avoid scheduling starvation.

[0077] After sorting the pending tasks, the computing device checks whether each task meets the resource constraints according to the queue order and the available resources, and selects one or more qualified tasks as target offline rendering tasks. Resource constraints can be time constraints and computational constraints. Time constraints mean that the predicted rendering completion time of a task does not exceed the total duration of a future preset period, ensuring that the task can be fully executed within a single scheduling cycle; computational constraints mean that the minimum GPU computing power required for task execution does not exceed the current remaining available resources, ensuring that the reserved computing power for online services is not encroached upon.

[0078] During the verification process, tasks that meet the constraints are included in the target offline rendering task set, and their occupied computing power is deducted simultaneously while the remaining available resources are updated. Tasks that do not meet the constraints are skipped directly until the task queue is traversed completely or the remaining available resources cannot support any task in the queue. If the available resources are 0, the computing device directly terminates the screening process, does not select any target tasks, and strictly adheres to the core principle of prioritizing online business computing power.

[0079] This application embodiment significantly reduces the computational overhead of task scheduling through a standardized sorting and filtering mechanism, improves filtering efficiency and scenario adaptability, and solves the problems of blind offline task scheduling, easy execution timeout, fragmented computing power, and difficulty in scheduling long-cycle tasks in the prior art. It not only ensures the service quality of online services, but also maximizes the use of idle computing power and significantly improves the overall utilization efficiency of the GPU resource pool.

[0080] S106: The computing device schedules the target offline rendering task to be executed within a future preset period.

[0081] The computing device obtains the target offline rendering task list from the output, synchronously retrieves the core parameters such as the required GPU computing power specifications and the predicted rendering completion time for each task, and allocates matching GPU resources to each target task from the idle computing power of the resource pool at the start node of the future preset period, creates the corresponding rendering running instance, deploys the adapted rendering engine environment, loads rendering materials and configuration parameters, starts the rendering process, and completes the scheduling and start of the target offline rendering task.

[0082] During the execution of the target offline rendering task, the computing device collects real-time data on the actual computing resource usage of the online cloud rendering service at a preset frequency through the resource pool monitoring system, and also collects data on the execution progress, resource usage, and running status of the offline rendering task. If the actual resource usage of the online service exceeds the previously predicted reserved resource amount, a fallback mechanism is immediately triggered. According to the preset shutdown priority, the currently executing offline rendering task is suspended or terminated to quickly release the occupied GPU computing resources, prioritizing the computing power supply and low-latency service quality of the online service. After the online service load falls back to the reserved resource range, the offline task rendering is resumed according to the task execution progress, or the scheduling and matching are re-completed. At the same time, in case of abnormal situations such as rendering process crashes or hardware failures, the computing device will automatically perform retry or rescheduling operations to ensure the integrity of the offline rendering task execution.

[0083] Once the target offline rendering task has completed all rendering computations, the computing device automatically completes the closed-loop processing of the entire scheduling process. The computing device collects the rendering output files, uploads them to a pre-set distributed storage system, generates a task execution report including actual execution time, resource consumption, and rendering result status, and pushes a completion notification to the user who submitted the task. Subsequently, the computing device immediately releases the GPU computing resources occupied by the task, cleans up the running instance and environment, reintegrates the released resources into the resource pool's idle computing power management, and updates the available resource data in the resource pool for task allocation in subsequent scheduling cycles or for sudden resource demands from online services.

[0084] This application embodiment achieves reliable execution of offline rendering tasks through standardized scheduling, real-time backup protection, and closed-loop resource management. It also eliminates the impact of offline tasks on the quality of online business services from the root, and maximizes the avoidance of idle computing power through timely resource release.

[0085] To further realize closed-loop management of GPU computing resources throughout the entire process, avoid invalid resource occupation after the completion of offline rendering tasks, and improve the computing power reuse efficiency and elastic scheduling capability of the resource pool, this application embodiment also adds a GPU node load monitoring and automatic resource release process after the target offline rendering task is scheduled to be executed within the future preset period.

[0086] In a possible implementation, after scheduling the target offline rendering task to be executed within a preset future period, the method further includes: The computing device monitors the load data of the GPU nodes executing the target offline rendering task. After the target offline rendering task is completed, the computing device releases the corresponding GPU node resources according to the load data.

[0087] After the target offline rendering task is started and executed, the computing device uses the monitoring system provided by the resource pool to monitor the load data of all GPU nodes executing this task throughout the cycle. The load data is a core indicator that can quantitatively characterize the running status, resource occupancy, and task execution progress of the GPU nodes, and specifically includes the real-time occupancy rate of GPU cores, the occupancy rate of video memory, the running status of the rendering process, the rendering frame progress of the task, the node computing power binding status, etc. During the monitoring execution process, the computing device continuously collects the load data of the corresponding GPU nodes at a preset time granularity, filters out outliers and validates the effectiveness of the original data, and simultaneously judges the execution status of the task in real time based on the load data, accurately identifying different stages such as normal rendering of the task, abnormal process, and rendering completion, providing an accurate trigger basis for subsequent resource release.

[0088] When the computing device determines that the target offline rendering task has been completed based on the continuously collected load data, it immediately triggers the corresponding GPU node resource release process. When the rendering process of the corresponding GPU node has been normally terminated, the occupancy rates of the GPU cores and video memory have dropped back to the idle threshold, all target frames of the task have been rendered and the result files have been archived and stored, it is determined that the task has been completed. After the task is completed, the computing device issues a standardized resource release instruction to the corresponding GPU node, terminates all remaining processes related to this offline task, clears the rendering running environment and temporary cache files, releases the resource binding between this GPU node and the task, adds it back to the available idle computing power list of the resource pool, and simultaneously updates the total available resource data of the resource pool.

[0089] For a batch rendering task executed in parallel by multiple GPU nodes, the computing device can, based on the load data of each node, achieve individual release when a single-node task is completed or unified release after all tasks are completed, flexibly adapting to different rendering scenarios.

[0090] The embodiment of this application forms the scheduling, execution, monitoring, and release of offline rendering tasks through a full-cycle load monitoring and automated resource release mechanism, solving the problem of idle waste of computing power caused by the failure to release resources in a timely manner after the task is completed in the prior art. By judging the task completion status based on the load data, it is possible to achieve immediate release of resources after the task is completed in advance, without waiting for the end of the preset future period, greatly shortening the idle cycle of computing power resources and further improving the overall utilization efficiency of the GPU resource pool.

[0091] To further clarify the complete business execution process of the hybrid scheduling method for online and offline cloud rendering tasks in this application, and to define the collaborative logic and technical implementation path of each interactive entity, this application embodiment also provides a detailed description of the entire execution process of this method in conjunction with the business interaction architecture.

[0092] like Figure 2 As shown, the core interactive entities in this embodiment are the user, the rendering platform, and the virtualization platform. The entire process uses the rendering platform as the core scheduling carrier to achieve full lifecycle management of offline rendering tasks and hybrid scheduling of online and offline services. The user first uploads rendering materials, specifies the task start time, and related configuration parameters through the rendering platform's interactive entry point. After receiving the offline rendering task submitted by the user, the rendering platform parses the rendering materials and extracts the corresponding feature parameters. Then, it calls the pre-trained first prediction model and, combined with the benchmark information of various GPU computing power in the GPU computing power information database, calculates the predicted rendering completion time required for the task to be executed on various GPUs, and simultaneously calculates the amount of computing power resources required for task execution. At the same time, based on the historical resource consumption data of online cloud rendering tasks, the rendering platform calculates the predicted resource amount for online services within a future preset period using a second prediction model, determines the available resource amount for offline task scheduling within that period, and finally, combines the predicted rendering completion time, required resource amount, and available resource amount of the offline task to select and determine the target offline rendering task.

[0093] After the rendering platform determines the target offline rendering task, it sends the scheduling instruction to the virtualization platform. The virtualization platform then allocates matching GPU node resources to the target task and executes the offline rendering task within a preset period. Once the task is completed, the virtualization platform feeds back the execution result to the rendering platform, which then displays the result to the user. Simultaneously, it releases the GPU node resources occupied by the task, updates the resource pool status, and forms a complete task scheduling loop. Ultimately, this maximizes the overall utilization efficiency of GPU computing resources while prioritizing the quality of online business services.

[0094] Based on the above method embodiments, this application also provides a scheduling device for online cloud rendering tasks and offline cloud rendering tasks, such as... Figure 3 As shown, the device includes: The acquisition module 301 is used to acquire task information of the offline rendering task to be processed, the task information including rendering materials; The first prediction module 302 is used to process the rendering material through the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processor GPU. The second prediction module 303 is used to predict the amount of resources required by the online cloud rendering task within a future preset period based on the historical resource consumption data of the online cloud rendering task and through the second prediction model. The determining module 304 is used to determine the amount of available resources that can be used to schedule offline rendering tasks within the future preset period based on the predicted resource amount and the total resource amount of the resource pool. The filtering module 305 is used to select a target offline rendering task from the offline rendering tasks to be processed based on the available resource quantity and the predicted rendering completion time. The scheduling module 306 is used to schedule the target offline rendering task to be executed within the future preset period.

[0095] In one possible implementation, the first prediction module 302 is specifically used to parse the rendering material to obtain at least one feature parameter of the rendering material; input the at least one feature parameter and the computing power performance benchmark information of the corresponding type of GPU in the GPU computing power information database into the first prediction model to obtain the predicted rendering completion time of the offline rendering task on the target type GPU.

[0096] In one possible implementation, the second prediction module 303 is specifically used to input the historical resource consumption data into the second prediction model to obtain the demand trend of GPU resources for online cloud rendering tasks in one or more future task cycles; and to determine the number of GPUs required for online cloud rendering tasks in the future preset cycle as the predicted resource quantity based on the demand trend.

[0097] In one possible implementation, the filtering module 305 is specifically used to sort the offline rendering tasks to be processed according to the rendering completion time of the offline rendering tasks on at least one type of GPU; and select one or more offline rendering tasks whose rendering completion time meets the resource constraints in sequence as the target offline rendering tasks based on the sorting results and the amount of available resources.

[0098] In one possible implementation, after the target offline rendering task is scheduled to be executed within the future preset period, the scheduling module 306 is further configured to monitor the load data of the GPU node executing the target offline rendering task; after the target offline rendering task is completed, the corresponding GPU node resources are released according to the load data.

[0099] In one possible implementation, the device further includes a construction module for constructing a GPU computing power information database; the GPU computing power information database includes computing power performance benchmark information corresponding to various types of GPUs supported by the resource pool.

[0100] In one possible implementation, the first prediction model is a task duration prediction model trained based on the Support Vector Regression (SVR) algorithm, and the second prediction model is an online resource prediction model trained based on the Prophet time series prediction algorithm.

[0101] This application also provides a control device. The control device may include a memory and a processor. The processor is used to execute the scheduling method for online cloud rendering tasks and offline cloud rendering tasks described in any of the above embodiments. The memory may be random access memory (RAM), flash memory, read-only memory (ROM), electronic programmable ROM (EPROM), registers, hard disk, removable disk, etc.

[0102] Memory can store computer instructions. When these instructions are executed by the processor, the processor can use them to implement scheduling methods for online and offline cloud rendering tasks. Memory can also store data.

[0103] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape) or a semiconductor medium (e.g., solid-state disk (SSD)).

[0104] This application also provides a readable storage medium for storing the methods provided in the above embodiments. For example, RAM, flash memory, ROM, EPROM, registers, hard disk, removable disk, or any other form of storage medium in the art.

[0105] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0106] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Regarding the methods disclosed in the embodiments, since they correspond to the product embodiments disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the description of the product embodiments.

[0107] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for scheduling online cloud rendering tasks and offline cloud rendering tasks, characterized in that, include: Obtain task information for offline rendering tasks to be processed, including rendering materials; The rendering material is processed by the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processing unit (GPU). Based on historical resource consumption data of online cloud rendering tasks, the predicted resource amount required by the online cloud rendering tasks in a future preset period is predicted by a second prediction model. Based on the predicted resource quantity and the total resource quantity of the resource pool, the available resource quantity that can be used to schedule offline rendering tasks within the future preset period is determined. Based on the available resources and the predicted rendering completion time, select a target offline rendering task from the offline rendering tasks to be processed. The target offline rendering task is scheduled to be executed within the preset future period.

2. The method according to claim 1, characterized in that, The step of processing the rendered material using a first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processing unit (GPU) includes: Analyze the rendered material to obtain at least one feature parameter of the rendered material; By inputting the at least one feature parameter and the benchmark information of the computing power of the corresponding type of GPU in the GPU computing power information database into the first prediction model, the predicted rendering completion time of the offline rendering task on the target type of GPU is obtained.

3. The method according to claim 1, characterized in that, The method of predicting the amount of resources required for the online cloud rendering task within a preset future period based on historical resource consumption data of the online cloud rendering task, using a second prediction model, includes: Input the historical resource consumption data into the second prediction model to obtain the demand trend of GPU resources for online cloud rendering tasks in one or more future task cycles; Based on the demand trend, the number of GPUs required for online cloud rendering tasks within the future preset period is determined as the predicted resource quantity.

4. The method according to claim 1, characterized in that, The step of selecting a target offline rendering task from the pending offline rendering tasks based on the available resource quantity and the predicted rendering completion time includes: The offline rendering tasks to be processed are sorted according to the rendering completion time of the offline rendering tasks on at least one type of GPU. Based on the sorting results and the available resources, one or more offline rendering tasks whose rendering completion time meets the resource constraints are selected in sequence as the target offline rendering tasks.

5. The method according to claim 1, characterized in that, After scheduling the target offline rendering task to be executed within the preset future period, the method further includes: Monitor the load data of the GPU nodes executing the target offline rendering task; After the target offline rendering task is completed, the corresponding GPU node resources are released based on the load data.

6. The method according to claim 1, characterized in that, The method further includes: Construct a GPU computing power information database; the GPU computing power information database includes computing power performance benchmark information corresponding to various types of GPUs supported by the resource pool.

7. The method according to claim 1, characterized in that, The first prediction model is a task duration prediction model trained based on the Support Vector Regression (SVR) algorithm, and the second prediction model is an online resource prediction model trained based on the Prophet time series prediction algorithm.

8. A scheduling device for online cloud rendering tasks and offline cloud rendering tasks, characterized in that, The device includes: The acquisition module is used to acquire task information of the offline rendering task to be processed, the task information including rendering materials; The first prediction module is used to process the rendering material through the first prediction model to obtain the predicted rendering completion time of the offline rendering task on at least one type of graphics processor GPU. The second prediction module is used to predict the amount of resources required by the online cloud rendering task within a future preset period based on the historical resource consumption data of the online cloud rendering task and through the second prediction model. The determination module is used to determine the amount of available resources that can be used to schedule offline rendering tasks within the future preset period based on the predicted resource amount and the total resource amount of the resource pool. The filtering module is used to select a target offline rendering task from the offline rendering tasks to be processed based on the available resource quantity and the predicted rendering completion time. The scheduling module is used to schedule the target offline rendering task to be executed within the preset future period.

9. A control device, characterized in that, It includes a processor and a memory, the memory being used to store programs, instructions, or code, and the processor being used to execute the programs, instructions, or code in the memory to complete the scheduling method for online cloud rendering tasks and offline cloud rendering tasks as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The system contains a computer program that is loaded by a processor to execute the scheduling method for online and offline cloud rendering tasks as described in any one of claims 1-7.

Images