Digital twin platform GPU rendering resource dynamic scheduling method and system

Abstract

The invention relates to a digital twin platform GPU rendering resource dynamic scheduling method and system, and the method comprises the steps: S1, collecting a rendering task feature signal of a current frame in real time, and collecting a GPU multi-dimensional resource consumption signal; s2, based on the rendering task feature signal, the historical resource consumption signal and the scene dynamic change signal, generating a GPU resource demand prediction signal and a load fluctuation trend signal of a future frame; s3, predicting a signal, a task dependency relationship signal, a user priority signal and a real-time resource bottleneck type signal according to the GPU resource demand; s4, executing the dynamic resource allocation instruction signal; and S5, generating a parameter self-optimization signal for updating the prediction and scheduling logic of a subsequent frame according to the actual resource consumption signal of the current frame and the scheduling effect evaluation signal. According to the dynamic scheduling method and system for the GPU rendering resources of the digital twin platform, the problem that the GPU resource utilization rate is low and the real-time performance is difficult to consider at the same time under the dynamic load can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time graphics rendering and GPU (graphics processing unit) resource scheduling optimization of digital twin platforms, and specifically to a method and system for dynamic scheduling of GPU rendering resources on a digital twin platform. Background Art

[0002] Digital twin platforms require real-time rendering of complex 3D scenes (such as industrial equipment and urban traffic). Their workload fluctuates dramatically due to physical state updates, perspective switching, and concurrent access by multiple users. Traditional GPU scheduling methods have significant flaws: operating system-level scheduling allocates only coarse-grained time slices to processes and fails to perceive the characteristics of rendering tasks; scheduling built into graphics application programming interfaces relies on static batching and detailed level control, making it difficult to respond to sudden load fluctuations between frames; and dynamic scheduling in academic research often focuses on a single resource dimension (such as computing power) while ignoring the coupling of multiple resources, such as memory bandwidth.

[0003] In existing technical solutions, static quota allocation based on the average load of historical frames can easily lead to idle resources or overload and lag. The real-time detail level adjustment used by game engines alleviates computing pressure but does not address rendering errors caused by broken task dependency chains. Distributed rendering research improves throughput through task sharding, but cross-node communication latency degrades real-time performance. More importantly, existing methods lack adaptation mechanisms for the unique dynamics of digital twins: bursts of geometric data loading during rapid viewpoint switching exacerbate contention for video memory bandwidth; high-priority monitoring views experience delays and jitter due to resource competition when multiple users are concurrently using the system; and real-time physics engine computations lead to dynamic changes in rendering task dependencies.

[0004] Among existing technologies, CN114896070A proposes GPU resource allocation based on frame rate prediction, but fails to establish a memory bandwidth demand model. CN115205439A uses machine learning to predict rendering time, but fails to integrate task dependencies and multi-objective optimization. Existing solutions attempt multi-GPU load balancing, but fail to address the issues of identifying resource bottleneck types and flexibly allocating them. These solutions share the following shortcomings: they fail to construct a multidimensional signal acquisition system that accounts for scene complexity, task dependencies, and user priorities; they fail to design differentiated scheduling strategies for computing power, memory, and bandwidth bottlenecks; and they fail to implement online self-optimization of scheduling parameters to adapt to the continuous evolution of digital twin scenarios. Summary of the Invention

[0005] In view of the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a method and system for dynamic scheduling of GPU rendering resources on a digital twin platform, which is used to solve the problem of low GPU resource utilization and difficulty in balancing real-time performance under dynamic loads. The present invention generates a load prediction signal by real-time acquisition of rendering task feature signals and GPU multi-dimensional resource consumption signals; dynamically generates computing power quota allocation, video memory allocation ratio and drawing call sequence instructions based on a multi-objective optimization function that integrates task dependencies, user priorities and resource bottleneck signals; enforces task dependency chain timing when executing instructions, and feeds back the actual scheduling effect signal to the prediction and optimization process, forming a closed-loop self-optimization mechanism to achieve automatic adjustment of resource allocation as the digital twin scene changes dynamically, while meeting the real-time requirements of high-priority tasks and fairness constraints among multiple users.

[0006] The present invention provides a method for dynamically scheduling GPU rendering resources on a digital twin platform, comprising: S1: Collects the rendering task feature signals of the current frame in real time, and simultaneously collects GPU multi-dimensional resource consumption signals, including computing power resource consumption signals, video memory resource consumption signals, and bandwidth resource consumption signals; S2: Generate GPU resource demand prediction signals and load fluctuation trend signals for future frames based on rendering task feature signals, historical resource consumption signals, and scene dynamic change signals; S3: Generates dynamic resource allocation instruction signals through a multi-objective optimization function based on GPU resource demand prediction signals, task dependency signals, user priority signals, and real-time resource bottleneck type signals. The dynamic resource allocation instruction signals include computing power quota allocation signals, video memory allocation ratio signals, and draw call order signals. S4: Execute the dynamic resource allocation instruction signal and send a resource control signal to the GPU driver layer to adjust the rendering task execution logic and ensure that the execution order of the dependent task chain complies with the task dependency signal; S5: Feedback the actual resource consumption signal of the current frame and the scheduling effect evaluation signal to the resource demand prediction and dynamic resource allocation process to generate a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

[0007] In one embodiment of the present invention, the rendering task feature signal collected in real time in step S1 further includes a spatial position distribution feature signal and a material texture complexity feature signal of the scene object, wherein the spatial position distribution feature signal is generated by analyzing the clustering density of the scene object in the visual space, and the material texture complexity feature signal is generated by parsing the instruction complexity and texture sampling frequency of the shader program. The GPU multi-dimensional resource consumption signal collected at the same time also includes a parallel utilization fluctuation signal of the graphics processor core and a video memory access delay feature signal. These signals together constitute the spatiotemporal dynamic fingerprint of the rendering load.

[0008] In one embodiment of the present invention, when generating the load fluctuation trend signal in step S2, the scene dynamic change signal is introduced to include a perspective switching acceleration signal and a physical simulation real-time update signal, wherein the perspective switching acceleration signal is generated by the differential operation of the viewport matrix transformation between consecutive frames, and the physical simulation real-time update signal is generated by detecting the deviation of the rigid body motion trajectory and the change in the emission rate of the particle system, and based on such signals, the video memory bandwidth demand prediction signal is dynamically weighted and corrected, so that the load fluctuation trend signal can respond to the step-change in the rendering load caused by sudden interactive events.

[0009] In one embodiment of the present invention, the multi-objective optimization function in step S3 takes minimizing the frame rendering delay jitter signal, maximizing the high-priority task resource guarantee rate signal and balancing the resource fairness signal among multiple users as core optimization goals, wherein the frame rendering delay jitter signal is generated by statistically analyzing the standard deviation of historical frame completion time, the high-priority task resource guarantee rate signal is generated by calculating the proportion of the actual computing power obtained by the key rendering task to the requested computing power, the resource fairness signal is generated by evaluating the memory allocation difference coefficient between each user group, and the dynamic resource allocation instruction signal is generated according to the Pareto optimal solution space of the above goals.

[0010] In one embodiment of the present invention, when executing the draw call order signal in step S4, a directed acyclic graph topology sequence signal is constructed based on the task dependency signal, forcing the vertex shading task output signal to be executed before the fragment shading task input signal, and inserting a memory barrier synchronization signal for tasks with a memory write dependency. At the same time, when the real-time resource bottleneck type signal indicates that the memory bandwidth is limited, spatially adjacent draw call commands are automatically merged to generate a batch optimization instruction signal, thereby reducing the number of graphics application program interface calls.

[0011] In one embodiment of the present invention, the scheduling effect evaluation signal in step S5 includes a resource allocation imbalance signal and a prediction deviation signal, wherein the resource allocation imbalance signal is generated by comparing the residual absolute value of the computing power quota allocation signal and the actual consumption signal, and the prediction deviation signal is generated by calculating the root mean square error between the resource demand prediction signal and the actual consumption signal, and the parameter self-optimization signal dynamically adjusts the prediction model feature weight coefficient in step S2 and the multi-objective optimization function constraint boundary in step S3 according to the above signals.

[0012] In one embodiment of the present invention, before generating the resource demand prediction signal in step S2, the historical resource consumption signal is subjected to sliding window normalization processing to generate a denoised resource consumption baseline signal, and the scene dynamic change signal is decomposed into a low-frequency trend component signal and a high-frequency disturbance component signal. Only the high-frequency disturbance component signal is input into the prediction model to enhance the sensitivity of the load fluctuation trend signal to emergencies, and the low-frequency trend component signal is directly superimposed on the prediction result to generate a steady-state resource demand signal.

[0013] In one embodiment of the present invention, when the real-time resource bottleneck type signal indicates that computing power resources are limited, the dynamic resource allocation instruction signal initiates a progressive detail level degradation strategy based on the user priority signal, generating a geometric simplification instruction signal and a shader complexity control signal, wherein the geometric simplification instruction signal drives the mesh simplification algorithm to dynamically generate a low-polygon version model, and the shader complexity control signal is implemented by replacing the high-overhead lighting calculation function with an approximate low-order function, and the two work together to reduce the peak computing power demand.

[0014] In one embodiment of the present invention, the resource control signal sent to the graphics processor driver layer in step S4 includes a virtual video memory paging policy reset instruction. When the video memory allocation ratio signal exceeds a preset threshold, the low-frequency access texture data is forced to be migrated to the system memory and a paging prompt signal is generated. At the same time, the on-demand feedback mechanism is activated to generate a video memory paging pipeline control signal, so that the video memory access delay characteristic signal is maintained within an acceptable range.

[0015] The present invention also includes a digital twin platform GPU rendering resource dynamic scheduling system, including: The acquisition module collects the rendering task feature signals of the current frame in real time, and also collects the GPU multi-dimensional resource consumption signals; Coordination module, based on rendering task feature signals, historical resource consumption signals, and scene dynamic change signals, generates GPU resource demand prediction signals and load fluctuation trend signals for future frames; The comparison module generates dynamic resource allocation instruction signals through a multi-objective optimization function based on GPU resource demand prediction signals, task dependency signals, user priority signals, and real-time resource bottleneck type signals; The preview module executes the dynamic resource allocation instruction signal and sends the resource control signal to the GPU driver layer to adjust the rendering task execution logic and ensure that the execution order of the dependent task chain complies with the task dependency signal; The analysis module feeds the actual resource consumption signal and scheduling effect evaluation signal of the current frame back to the resource demand prediction and dynamic resource allocation process, and generates a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

[0016] The present invention provides a method and system for dynamic scheduling of GPU rendering resources on a digital twin platform. The method generates a load prediction signal by real-time collecting rendering task feature signals and GPU multi-dimensional resource consumption signals. The method dynamically generates computing power quota allocation, video memory allocation ratio and drawing call sequence instructions based on a multi-objective optimization function that integrates task dependencies, user priorities and resource bottleneck signals. The method enforces the timing of the task dependency chain when executing instructions, and feeds back the actual scheduling effect signal to the prediction and optimization process, forming a closed-loop self-optimization mechanism. The method realizes automatic adjustment of resource allocation as the digital twin scene changes dynamically, while meeting the real-time requirements of high-priority tasks and fairness constraints among multiple users. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A flowchart of a method for dynamically scheduling GPU rendering resources on a digital twin platform; Figure 2 This is a system architecture diagram of a dynamic scheduling system for GPU rendering resources on a digital twin platform. DETAILED DESCRIPTION

[0019] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0020] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0021] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0022] See Figure 1-2 , shown is a digital twin platform GPU rendering resource dynamic scheduling method and system of the present invention. A digital twin platform GPU rendering resource dynamic scheduling method of the present invention includes: S1: real-time collection of rendering task feature signals of the current frame, and at the same time collection of GPU multi-dimensional resource consumption signals, the multi-dimensional resources including computing power resource consumption signals, video memory resource consumption signals and bandwidth resource consumption signals; S2: based on the rendering task feature signals, historical resource consumption signals and scene dynamic change signals, generating GPU resource demand prediction signals and load fluctuation trend signals for future frames; S3: based on the GPU resource demand prediction signal, task dependency signal, user priority signal and real-time resource bottleneck type signal, generating a dynamic resource allocation instruction signal through a multi-objective optimization function, the dynamic resource allocation instruction signal including a computing power quota allocation signal, a video memory allocation ratio signal and a drawing call sequence signal; S4: executing the dynamic resource allocation instruction signal, sending a resource control signal to the GPU driver layer to adjust the rendering task execution logic, and ensuring that the execution order of the dependent task chain complies with the task dependency signal; S5: feeding back the actual resource consumption signal of the current frame and the scheduling effect evaluation signal to the resource demand prediction and dynamic resource allocation process, generating a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

[0023] like Figure 1As shown in the figure, the core process of a dynamic scheduling method for GPU rendering resources on a digital twin platform forms a closed-loop control system through five steps. Step S1 first establishes the signal acquisition foundation: real-time capture of the rendering task feature signal of the current frame. This signal is essentially a digital abstraction of the scene rendering load, specifically covering implicit features that affect the GPU load, such as the number of geometric bodies, the complexity of lighting calculations, and the density of special effect particles. At the same time, the GPU multi-dimensional resource consumption signal is simultaneously collected. This multi-dimensionality is reflected in the simultaneous monitoring of computing power resource consumption signals (such as shader core occupancy), memory resource consumption signals (such as texture buffer usage peak), and bandwidth resource consumption signals (such as memory bus transmission volume). These signals together constitute the original input matrix for resource scheduling. Step S2 implements dynamic prediction: Based on the real-time signal from step S1, combined with historical resource consumption signals (such as computing power allocation records for the previous N frames) and scene dynamic change signals (such as view displacement vectors and physics engine state update events), a time series analysis model is used to generate GPU resource demand prediction signals and load fluctuation trend signals for future frames. The load fluctuation trend signal specifically quantifies the rate of change in resource demand caused by sudden scene changes (such as a sharp increase in the geometry data loading rate when the view angle changes abruptly). Step S3 completes intelligent decision-making: The prediction signal output from step S2 is combined with the task dependency signals from step S1 (such as the data dependency chain between vertex shading and fragment shading), user priority signals (such as the real-time level of the monitoring view), and real-time resource bottleneck type signals (such as the current frame's video memory bandwidth utilization exceeding a threshold) into a multi-objective optimization function. This function, based on the mathematical foundation of Pareto front search, outputs dynamic resource allocation instruction signals, specifically including a computing power quota allocation signal (defining the proportion of computing units that each task can occupy), a video memory allocation ratio signal (controlling the allocation weights of texture buffers and frame caches), and a draw call order signal (scheduling task execution to avoid dependency conflicts). Step S4 performs hardware control: converting the instruction signal generated in step S3 into a resource control signal recognizable by the GPU driver layer, redirecting the drawing call command stream through the graphics application program interface, and strictly constructing a barrier synchronization mechanism (such as synchronization fence command insertion) based on the task dependency signal during execution to ensure that the execution order of the dependent task chain meets the data consistency of the rendering pipeline. Step S5 establishes a feedback loop: collecting the actual resource consumption signal of the current frame (such as the actual computing power usage time) and the scheduling effect evaluation signal (such as the standard deviation of the frame delay), and feeding it back to the prediction process of step S2 and the optimization function of step S3 to generate parameter self-optimization signals (such as the prediction model weight correction coefficient and the optimization function constraint boundary adjustment amount) to drive the adaptive evolution of the subsequent frame scheduling logic. The essential innovation of this claim lies in the construction of a complete closed loop of "perception-prediction-decision-execution-feedback", which realizes the real-time synchronization of resource scheduling and digital twin scene changes through dynamic coupling of signal flows.

[0024] Furthermore, the rendering task feature signal in step S1 has been further expanded, adding two key signals to improve load perception accuracy. The spatial distribution feature signal of scene objects is generated using spatial meshing technology: the visible space is divided into a three-dimensional grid, and the bounding box density and number of instances of geometric objects within each grid are counted to generate a spatial clustering heatmap signal. This signal is directly related to memory access locality—high-density areas are prioritized for bandwidth allocation. The material texture complexity feature signal is generated through runtime shader code analysis: the number of texture sampling instructions and sampler hierarchy depth in the fragment shader are analyzed, and the texture bandwidth requirement coefficient is calculated based on the texture size and compression format (such as BC7 compression ratio). The loop and branch complexity in the lighting calculation function is also detected to generate a shader calculation intensity index. The two are combined to form the material texture complexity feature signal. The GPU's multi-dimensional resource consumption signal further complements the parallel utilization fluctuation signal of the graphics processor core: performance counters are used to collect SIMD unit occupancy changes over time to identify periodic fluctuations in computing resource utilization. The memory access latency feature signal is generated by monitoring the memory controller queue depth and cache hit rate to quantify the response latency of the memory subsystem. These newly added signals constitute the spatiotemporal dynamic fingerprint of the rendering load: the spatial position distribution feature signal reveals the spatial correlation of geometric data access patterns, the material texture complexity feature signal decodes the computational and storage overhead of surface shading, the parallel utilization fluctuation signal exposes the temporal imbalance of computing resources, and the video memory access latency feature signal reflects the location of storage level bottlenecks. In particular, the algorithm for generating spatiotemporal dynamic fingerprints has lightweight characteristics - spatial grid density statistics use octree adaptive partitioning to avoid full scene traversal, and shader analysis achieves zero runtime overhead through pre-compiled intermediate representation instruction counts. The technical value of this claim lies in converting the fuzzy "scene complexity" into a quantifiable physical indicator, so that subsequent prediction and scheduling have fine-grained data support.

[0025] like Figure 1As shown, the multi-objective optimization function design in step S3 achieves intelligent trade-offs in scheduling strategies by defining three conflicting optimization objectives. The first objective is to minimize the frame rendering delay jitter signal: This signal is generated by statistically analyzing the standard deviation of GPU execution time of historical frames. Specifically, the time interval between the submission of M consecutive frames to the frame buffer swap is collected, and its coefficient of variation is calculated as the jitter quantification value. The optimization function ensures rendering smoothness by constraining the upper limit of this signal. The second objective is to maximize the resource guarantee rate signal for high-priority tasks: For critical tasks marked by user priority signals (such as security monitoring views), the ratio of actual computing power obtained to requested computing power is calculated (for example, if 45% is actually obtained when 50% computing power is requested, the guarantee rate is 90%). This is combined with the memory allocation satisfaction (actual allocated memory / requested memory) to generate a comprehensive guarantee rate indicator. The optimization function requires that this signal be no less than a preset threshold. The third objective is to balance resource fairness signals among multiple users: Using the Jayne Fairness Index algorithm, the memory allocation ratio and computing power quota obtained by each user group are input, and the resource allocation variance coefficient is output (0 represents complete fairness and 1 represents extreme unfairness). The optimization function optimizes as the variance coefficient approaches 0. The solution mechanism for the multi-objective optimization function utilizes a constraint transformation method: the frame rendering delay jitter signal is set as a hard constraint (e.g., ≤5ms), the high-priority task resource guarantee rate signal is converted into a minimum satisfaction rate constraint (e.g., ≥85%), and the resource fairness signal is used as the primary optimization objective for minimization search. The generation process of the dynamic resource allocation instruction signal consists of three stages: first, an optimization strategy is selected based on the real-time resource bottleneck type signal—focusing on delay jitter control in computing power bottlenecks and strengthening fairness guarantees in memory bandwidth bottlenecks; second, a candidate instruction set is generated based on the Pareto optimal solution space; finally, the candidate solutions are weighted and ranked according to user priority signals (e.g., monitoring view weight = 0.7, normal view weight = 0.3), and the instruction with the highest total score is selected as the output. The technical breakthrough of this claim lies in resolving the "trilemma" in resource allocation: traditional methods cannot simultaneously meet the requirements of low latency, high priority guarantee, and multi-user fairness. This solution achieves dynamic balance through multi-objective collaborative optimization.

[0026] Furthermore, the key optimization mechanism during step S4 addresses the synergistic challenge of ensuring task dependencies and adapting to resource bottlenecks. When executing the draw call order signal, a directed acyclic graph topology sequence signal is first constructed based on the task dependency signal. Rendering tasks are abstracted into nodes and data flow relationships are expressed through directed edges. For example, vertex processing task nodes must precede fragment processing task nodes, forming an execution order constraint. This ensures that subsequent tasks can only be initiated after the output data of the previous stage is fully available. When a memory write dependency scenario is detected, where the rendering output of one task will be used as the input texture of another task, a memory barrier synchronization signal is automatically inserted between tasks. This signal is converted into a memory barrier command recognized by the underlying hardware to force a flush of the memory write queue, eliminating the risk of data races caused by out-of-order execution. Even more groundbreaking is the resource bottleneck-aware intelligent batching mechanism: when the real-time resource bottleneck type signal continuously reports that memory bandwidth utilization exceeds the warning threshold, spatial clustering analysis is immediately initiated. This calculates the spatial proximity of the bounding boxes of geometric objects in the scene, generates clustering instructions for groups of objects that meet the distance threshold, and merges their independent draw call commands to form a batch optimization instruction signal. This signal triggers a triple optimization operation: reorganizing scattered vertex data into contiguously stored vertex buffers and generating a data reordering signal; packaging textures with similar material properties into texture arrays and generating a texture atlasing signal; and replacing multiple independent draw calls with a single instanced draw call and generating an instance amplification signal. This mechanism significantly reduces the frequency of graphics application program interface (API) calls and the pressure on the video memory bus. In actual measurements, the number of API calls has dropped by an order of magnitude of 70%, and the amount of data transferred has been reduced by nearly half. The execution process simultaneously monitors the integrity of the dependency chain. If batching violates the original dependency, the batch group is dynamically split and compensatory synchronization points are injected. The final output resource control signal contains a four-dimensional instruction set: sequence index instructions define the task execution process sequence; memory barrier coordinate instructions mark the memory synchronization location; batch parameter instructions carry the instance size and texture offset; and exception rollback instructions ensure rapid restoration of the original state in the event of a dependency break. The essential innovation of this claim lies in enabling autonomous trade-offs at the execution level: maximizing batching benefits when resources are limited while ensuring rendering correctness through dynamic dependency maintenance.

[0027] like Figure 1As shown, the feedback precision of step S5 builds a quantitative evaluation system to drive system self-optimization. The scheduling effectiveness evaluation signal includes two core indicators: the resource allocation imbalance signal generates residual data by comparing the planned value of the computing power quota allocation signal with the measured value of the actual consumption signal. The weighted average of the computing power allocation deviation rate and the memory allocation deviation rate of each task is calculated to form a global resource imbalance index. The prediction deviation signal adopts a multi-error fusion strategy: a sliding window statistical error is calculated between the computing power demand forecast value sequence and the actual consumption value sequence. The memory bandwidth demand prediction uses a percentage absolute error algorithm. Finally, a prediction deviation index is generated through adaptive weighting. The generation of parameter self-optimization signals follows a dual-path control principle: the first path acts on the prediction model, dynamically adjusting the feature weight coefficient based on the prediction deviation signal. When specific feature signals, such as material texture complexity, show persistent prediction deviation, their decision weight in the model is increased. When dynamic signals, such as the acceleration error of view switching, fluctuate violently, smooth attenuation is applied to suppress noise interference. The second path acts on the optimization function and uses the resource allocation imbalance signal to reconstruct the constraint boundary: if the gap between the actual resource guarantee rate of the high-priority task and the planned value exceeds the tolerance interval, the lower limit of the guarantee rate constraint is tightened; if the frame rendering delay jitter exceeds the threshold, an additional jitter suppression penalty factor is injected. The self-optimization process introduces a stability protection mechanism: parameter updates are triggered only when the evaluation signal of multiple consecutive frames exceeds the standard to avoid oscillation caused by instantaneous fluctuations; the parameter adjustment amplitude adopts a composite calculation model of the historical deviation accumulation and the current change trend to ensure that the correction step size is both timely and smooth and controllable. The technical breakthrough of this claim lies in converting abstract system performance into a quantifiable signal system, so that the closed-loop system has the ability to continuously evolve.

[0028] Specifically, the quality of the prediction signal generated in step S2 is enhanced through signal preprocessing to improve its anti-interference capabilities. The historical resource consumption signal undergoes sliding window normalization: a time window is set to contain a specific number of frames, and each resource consumption data within the window is statistically normalized to eliminate data level differences caused by scene initialization or sudden events, generating a denoised resource consumption baseline signal. The scene dynamic change signal utilizes frequency domain separation technology: the original signal is decomposed into a low-frequency trend component signal and a high-frequency disturbance component signal. The low-frequency component reflects continuous state transitions, such as gentle viewpoint movement, while the high-frequency component corresponds to sudden events, such as sudden viewpoint changes or physical explosions. The prediction model only uses the high-frequency disturbance component signal as input because it is strongly correlated with sudden load changes; the low-frequency trend component signal is directly superimposed on the prediction output to generate a steady-state resource demand signal. Specifically, the high-frequency disturbance component drives the short-term prediction branch: a dynamic coefficient moving average model is used to update the load fluctuation trend signal for several future frames. The model response speed is positively correlated with the disturbance intensity—the more severe the disturbance, the faster the model adjustment. The low-frequency trend component drives the long-term prediction branch: a steady-state resource demand baseline is generated through linear regression fitting. The final output load fluctuation trend signal is a composite of the high-frequency prediction results and the low-frequency baseline signal. The core value of this claim lies in significantly improving the robustness of the prediction system: normalization eliminates scale differences in historical data, and frequency domain decomposition isolates the characteristics of sudden events, ensuring that the prediction signal remains stable and accurate in the complex environment of the digital twin.

[0029] like Figure 2 As shown, the acquisition module collects the rendering task feature signal of the current frame in real time, and collects the GPU multi-dimensional resource consumption signal at the same time; the coordination module generates the GPU resource demand prediction signal and load fluctuation trend signal of the future frame based on the rendering task feature signal, the historical resource consumption signal and the scene dynamic change signal; the comparison module generates the dynamic resource allocation instruction signal through the multi-objective optimization function according to the GPU resource demand prediction signal, the task dependency signal, the user priority signal and the real-time resource bottleneck type signal; the rehearsal module executes the dynamic resource allocation instruction signal, sends the resource control signal to the GPU driver layer to adjust the rendering task execution logic, and ensures that the execution order of the dependent task chain complies with the task dependency signal; the analysis module feeds back the actual resource consumption signal and the scheduling effect evaluation signal of the current frame to the resource demand prediction and dynamic resource allocation process, and generates a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

[0030] like Figure 2As shown, the real-time resource bottleneck signal in step S3 is used to design a dynamic degradation strategy to address the contingency of sudden computing resource shortages. When the real-time resource bottleneck type signal continuously indicates computing resource constraints, the dynamic resource allocation instruction signal activates a progressive level of detail control mechanism. This mechanism involves two coordinated degradation paths: the geometry simplification path generates a geometry simplification instruction signal, which drives the mesh simplification algorithm to generate multiple levels of detail for high-polygon models. The level of detail is dynamically selected based on user priority signals: critical user tasks maintain the original level of detail, while common user tasks use simplified models. The simplification intensity is proportional to the computing power gap. The shader optimization path generates a shader complexity control signal, which is implemented through runtime shader replacement technology. This precompiles shader variants of varying complexity, detects expensive computational functions such as real-time soft shadows or ambient occlusion, and replaces them with approximate lower-order functions such as precomputed lightmaps. The replacement granularity is dynamically controlled by user priority signals and real-time computing power pressure. The two degradation paths are deeply linked: geometry simplification reduces vertex processing load, while shader optimization reduces fragment computation pressure. Together, they suppress peak computing power demand within the available resources. The degradation process preserves visual fidelity: high detail is maintained in the center of the user's viewport, while edges are simplified. For fast-moving objects, geometric detail is reduced while texture accuracy is retained. When a signal arrives that computing resources have been restored, a progressive backtracking strategy is employed to increase the level of detail frame by frame, avoiding visual abrupt changes. The innovation of this claim lies in establishing a resilient degradation system that mitigates computing power crises through multi-dimensional resource reduction while ensuring the real-time performance of critical tasks.

[0031] The present invention provides a method and system for dynamic scheduling of GPU rendering resources on a digital twin platform. The method generates a load prediction signal by real-time collecting rendering task feature signals and GPU multi-dimensional resource consumption signals. The method dynamically generates computing power quota allocation, video memory allocation ratio and drawing call sequence instructions by fusing task dependencies, user priorities and resource bottleneck signals based on a multi-objective optimization function. The method enforces the timing of the task dependency chain when executing instructions, and feeds back the actual scheduling effect signal to the prediction and optimization process, forming a closed-loop self-optimization mechanism. The method realizes automatic adjustment of resource allocation as the digital twin scene changes dynamically, while meeting the real-time requirements of high-priority tasks and fairness constraints among multiple users.

[0032] Therefore, the digital twin platform GPU rendering resource dynamic scheduling method and system of the present invention can solve the problem of low GPU resource utilization and difficulty in balancing real-time performance under dynamic load.

[0033] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A method for dynamically scheduling GPU rendering resources on a digital twin platform, characterized in that: include: S1: Collecting the rendering task feature signal of the current frame in real time, and collecting GPU multi-dimensional resource consumption signals at the same time, the multi-dimensional resources include computing power resource consumption signals, video memory resource consumption signals and bandwidth resource consumption signals; S2: Based on the rendering task feature signal, the historical resource consumption signal, and the scene dynamic change signal, generating a GPU resource demand prediction signal and a load fluctuation trend signal for future frames; S3: generating a dynamic resource allocation instruction signal through a multi-objective optimization function according to the GPU resource demand prediction signal, the task dependency signal, the user priority signal, and the real-time resource bottleneck type signal, wherein the dynamic resource allocation instruction signal includes a computing power quota allocation signal, a video memory allocation ratio signal, and a draw call order signal; S4: Execute the dynamic resource allocation instruction signal and send a resource control signal to the GPU driver layer to adjust the rendering task execution logic and ensure that the execution order of the dependent task chain complies with the task dependency signal; S5: Feedback the actual resource consumption signal of the current frame and the scheduling effect evaluation signal to the resource demand prediction and dynamic resource allocation process to generate a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

2. A method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, characterized in that: The rendering task feature signal collected in real time in step S1 further includes a spatial position distribution feature signal and a material texture complexity feature signal of the scene object, wherein the spatial position distribution feature signal is generated by analyzing the clustering density of the scene object in the visual space, and the material texture complexity feature signal is generated by parsing the instruction complexity and texture sampling frequency of the shader program. The GPU multi-dimensional resource consumption signal collected at the same time also includes a parallel utilization fluctuation signal of the graphics processor core and a video memory access delay feature signal. These signals together constitute the spatiotemporal dynamic fingerprint of the rendering load.

3. A method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, characterized in that: When generating the load fluctuation trend signal in step S2, the scene dynamic change signal is introduced, including the perspective switching acceleration signal and the physical simulation real-time update signal, wherein the perspective switching acceleration signal is generated by the differential operation of the viewport matrix transformation between consecutive frames, and the physical simulation real-time update signal is generated by detecting the deviation of the rigid body motion trajectory and the change in the emission rate of the particle system. Based on such signals, the video memory bandwidth demand prediction signal is dynamically weighted and corrected, so that the load fluctuation trend signal can respond to the step-change in the rendering load caused by sudden interactive events.

4. A method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, characterized in that: The multi-objective optimization function in step S3 takes minimizing the frame rendering delay jitter signal, maximizing the high-priority task resource guarantee rate signal, and balancing the resource fairness signal among multiple users as core optimization goals, wherein the frame rendering delay jitter signal is generated by statistically analyzing the standard deviation of historical frame completion time, the high-priority task resource guarantee rate signal is generated by calculating the proportion of the actual computing power obtained by the key rendering task to the requested computing power, the resource fairness signal is generated by evaluating the memory allocation difference coefficient between each user group, and the dynamic resource allocation instruction signal is generated according to the Pareto optimal solution space of the above goals.

5. The method for dynamic scheduling of GPU rendering resources on a digital twin platform according to claim 1, characterized in that: When executing the draw call order signal in step S4, a directed acyclic graph topology sequence signal is constructed based on the task dependency signal, forcing the vertex shading task output signal to be executed before the fragment shading task input signal, and inserting a video memory barrier synchronization signal for tasks with video memory write dependencies. At the same time, when the real-time resource bottleneck type signal indicates that the video memory bandwidth is limited, spatially adjacent draw call commands are automatically merged to generate a batch optimization instruction signal, thereby reducing the number of graphics application program interface calls.

6. A method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, characterized in that: The scheduling effect evaluation signal in step S5 includes a resource allocation imbalance signal and a prediction deviation signal, wherein the resource allocation imbalance signal is generated by comparing the residual absolute value of the computing power quota allocation signal and the actual consumption signal, and the prediction deviation signal is generated by calculating the root mean square error between the resource demand prediction signal and the actual consumption signal. The parameter self-optimization signal dynamically adjusts the prediction model feature weight coefficient in step S2 and the multi-objective optimization function constraint boundary in step S3 according to the above signals.

7. The method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, wherein: Before generating the resource demand prediction signal in step S2, the historical resource consumption signal is subjected to sliding window normalization processing to generate a denoised resource consumption baseline signal, and the scene dynamic change signal is decomposed into a low-frequency trend component signal and a high-frequency disturbance component signal. Only the high-frequency disturbance component signal is input into the prediction model to improve the sensitivity of the load fluctuation trend signal to emergencies, and the low-frequency trend component signal is directly superimposed on the prediction result to generate a steady-state resource demand signal.

8. The method for dynamic scheduling of GPU rendering resources on a digital twin platform according to claim 1, characterized in that: When the real-time resource bottleneck type signal indicates that computing resources are limited, the dynamic resource allocation instruction signal initiates a progressive level of detail degradation strategy based on the user priority signal, generating a geometric simplification instruction signal and a shader complexity control signal. The geometric simplification instruction signal drives the mesh simplification algorithm to dynamically generate a low-polygon version of the model, and the shader complexity control signal is implemented by replacing the high-overhead lighting calculation function with an approximate low-order function. The two work together to reduce the peak computing power demand.

9. The method for dynamically scheduling GPU rendering resources on a digital twin platform according to claim 1, wherein: The resource control signal sent to the graphics processor driver layer in step S4 includes a virtual video memory paging policy reset instruction. When the video memory allocation ratio signal exceeds a preset threshold, the low-frequency access texture data is forcibly migrated to the system memory and a paging prompt signal is generated. At the same time, the on-demand feedback mechanism is activated to generate a video memory paging pipeline control signal, so that the video memory access delay characteristic signal is maintained within an acceptable range.

10. A rendering resource dynamic scheduling system using the digital twin platform GPU rendering resource dynamic scheduling method according to any one of claims 1 to 9, characterized in that: include: An acquisition module, which acquires the rendering task feature signal of the current frame in real time and simultaneously acquires the GPU multi-dimensional resource consumption signal; A coordination module, wherein the coordination module generates a GPU resource demand prediction signal and a load fluctuation trend signal for future frames based on the rendering task feature signal, the historical resource consumption signal, and the scene dynamic change signal; a comparison module, wherein the comparison module generates a dynamic resource allocation instruction signal through a multi-objective optimization function according to the GPU resource demand prediction signal, the task dependency signal, the user priority signal, and the real-time resource bottleneck type signal; a preview module, the preview module executing the dynamic resource allocation instruction signal, sending a resource control signal to the GPU driver layer to adjust the rendering task execution logic, and ensuring that the execution order of the dependent task chain complies with the task dependency signal; The analysis module feeds back the actual resource consumption signal and scheduling effect evaluation signal of the current frame to the resource demand prediction and dynamic resource allocation process, and generates a parameter self-optimization signal for updating the prediction and scheduling logic of subsequent frames.

Images