A digital twin platform GPU rendering resource dynamic scheduling method and system

Abstract

The application discloses a digital twin platform GPU rendering resource dynamic scheduling method and system, the method comprises the following steps: S1, real-time acquisition of the rendering task characteristic signal of the current frame, and simultaneously collecting the multi-dimensional resource consumption signal of the GPU; S2, based on the rendering task characteristic signal, the historical resource consumption signal and the scene dynamic change signal, the GPU resource demand prediction signal and the load fluctuation trend signal of the future frame are generated; S3, according to the GPU resource demand prediction signal, the task dependency relationship signal, the user priority signal and the real-time resource bottleneck type signal; S4, the dynamic resource allocation instruction signal is executed; S5, the actual resource consumption signal of the current frame and the scheduling effect evaluation signal are used to generate the parameter self-optimization signal for updating the prediction and scheduling logic of the subsequent frame. The application discloses a digital twin platform GPU rendering resource dynamic scheduling method and system, which can solve the problem that the GPU resource utilization is low and the real-time performance is difficult to be considered under dynamic load.

Description

TECHNICAL FIELD

[0001] The application relates to the field of real-time graphics rendering and GPU (graphics processing unit) resource scheduling optimization of a digital twin platform, and in particular to a digital twin platform GPU rendering resource dynamic scheduling method and system. BACKGROUND

[0002] A digital twin platform needs to render complex three-dimensional scenes (such as industrial equipment and urban traffic) in real time, and the load thereof fluctuates dramatically due to physical state updates, view angle switching and multi-user concurrent access. Traditional GPU scheduling methods have significant defects: operating system-level scheduling only allocates process time slices in a coarse-grained manner and cannot perceive the characteristics of rendering tasks; built-in scheduling of a graphics application programming interface relies on static batch processing and detail level control and is difficult to respond to inter-frame load mutations; dynamic scheduling in academic research focuses on a single resource dimension (such as computing power) and ignores the coupling of multiple resource dimensions such as memory bandwidth.

[0003] In the prior art, static quota allocation based on historical frame average load is prone to cause resource idling or overload lag; although instant detail level adjustment adopted by a game engine alleviates computing power pressure, it does not solve rendering errors caused by the breaking of a task dependency chain; distributed rendering research improves throughput through task fragmentation, but cross-node communication delay worsens real-time performance. More importantly, existing methods lack adaptive mechanisms for the dynamics specific to digital twins: sudden geometry data loading when the view angle is switched at a high speed intensifies memory bandwidth contention, delay jitter of high-priority monitoring views occurs due to resource competition when multiple users are concurrently accessing, and real-time operation of a physical engine causes dynamic changes in the dependency relationship of rendering tasks.

[0004] CN114896070A in the prior art proposes GPU resource allocation based on frame rate prediction, but does not establish a memory bandwidth demand model; CN115205439A uses machine learning to predict rendering time, but does not integrate task dependency and multi-objective optimization; existing technical solutions attempt to balance the load of multiple GPUs, but do not solve the problems of resource bottleneck type identification and elastic allocation. The common defects of these solutions are: a multi-dimensional signal collection system covering scene complexity, task dependency and user priority is not constructed; differentiated scheduling strategies for computing power, memory and bandwidth bottlenecks are not designed; and scheduling parameters are not optimized online to adapt to the continuous evolution of digital twin scenes. SUMMARY

[0005] In view of the above prior art disadvantages, the purpose of the present application is to provide a digital twin platform GPU rendering resource dynamic scheduling method and system for solving the problems of low GPU resource utilization and difficulty in balancing real-time performance under dynamic load. The present application generates a load prediction signal by real-time collection of rendering task characteristic signals and GPU multi-dimensional resource consumption signals; based on a multi-objective optimization function, the task dependency relationship, user priority and resource bottleneck signal are fused to dynamically generate a computing power quota allocation, a video memory allocation ratio and a drawing call order instruction; the task dependency chain timing is forcibly guaranteed when the instruction is executed, and the actual scheduling effect signal is fed back to the prediction and optimization process to form a closed-loop self-optimization mechanism, realizing automatic adjustment of resource allocation with dynamic changes in the digital twin scene, while meeting the real-time performance requirements of high-priority tasks and fairness constraints between multiple users.

[0006] The present application provides a digital twin platform GPU rendering resource dynamic scheduling method, comprising:

[0007] S1: real-time collection of rendering task characteristic signals of the current frame, and simultaneous collection of GPU multi-dimensional resource consumption signals, including computing power resource consumption signals, video memory resource consumption signals and bandwidth resource consumption signals;

[0008] S2: based on the rendering task characteristic signals, historical resource consumption signals and scene dynamic change signals, generating GPU resource demand prediction signals and load fluctuation trend signals for future frames;

[0009] S3: according to the GPU resource demand prediction signal, the task dependency relationship signal, the user priority signal and the 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 order signal;

[0010] 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 conforms to the task dependency relationship signal;

[0011] 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 to generate a parameter self-optimization signal for updating the prediction and scheduling logic of the subsequent frame.

[0012] In an embodiment of the present application, the rendering task feature signal collected in real time in step S1 further includes a spatial position distribution feature signal of the scene object and a material texture complexity feature signal, wherein the spatial position distribution feature signal is generated by analyzing the aggregation density of the scene object in the visual space, and the material texture complexity feature signal is generated by analyzing the instruction complexity and texture sampling frequency of the shader program; the GPU multi-dimensional resource consumption signal collected at the same time further includes a parallel utilization rate fluctuation signal of the graphics processor core and a memory access delay feature signal, and these signals together constitute the time-space dynamic fingerprint of the rendering load.

[0013] In an embodiment of the present application, when generating the load fluctuation trend signal in step S2, the scene dynamic change signal includes a view angle switching acceleration signal and a physical simulation real-time update amount signal, wherein the view angle switching acceleration signal is generated by differential operation of the matrix transformation of the view port between consecutive frames, and the physical simulation real-time update amount signal is generated by detecting the deviation of the rigid body motion trajectory and the change of the particle system emission rate, and the memory bandwidth demand prediction signal is dynamically weighted and corrected based on such signals, so that the load fluctuation trend signal can respond to the step mutation of the rendering load caused by the sudden interactive event.

[0014] In an embodiment of the present application, 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 the core optimization targets, wherein the frame rendering delay jitter signal is generated by calculating the standard deviation of the historical frame completion time, the high-priority task resource guarantee rate signal is generated by calculating the ratio of the actual obtained computing power to the applied computing power of the key rendering task, and the resource fairness signal is generated by evaluating the memory allocation difference coefficient between user groups, and the dynamic resource allocation instruction signal is generated according to the Pareto optimal solution space of the above targets.

[0015] In an embodiment of the present application, when executing the drawing call order signal in step S4, a directed acyclic graph topology sequence signal is constructed according to the task dependency relationship signal, it is forced to require that the vertex shading task output signal is executed before the fragment shading task input signal, and a memory barrier synchronization signal is inserted for the task with memory write dependence, and when the real-time resource bottleneck type signal indicates that the memory bandwidth is limited, the batch optimization instruction signal is automatically generated by merging the spatially adjacent drawing call commands, so as to reduce the number of graphics application program interface calls.

[0016] In an embodiment of the present application, the scheduling effect evaluation signal in step S5 includes a resource allocation imbalance degree signal and a prediction deviation degree signal, wherein the resource allocation imbalance degree signal is generated by comparing the absolute value of the residual of the computing power quota allocation signal and the actual consumption signal, the prediction deviation degree signal is generated by calculating the root mean square error of 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.

[0017] In an embodiment of the present application, 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 sudden events, and the low-frequency trend component signal is directly superimposed on the prediction result to generate a steady-state resource demand signal.

[0018] In an embodiment of the present application, when the real-time resource bottleneck type signal indicates that the computing power resource is limited, the dynamic resource allocation instruction signal starts a gradual detail level degradation strategy based on the user priority signal, generates a geometric simplification instruction signal and a shader complexity control signal, wherein the geometric simplification instruction signal drives the grid simplification algorithm to dynamically generate a low-polygon version model, and the shader complexity control signal is realized by replacing the high-overhead lighting calculation function with an approximate low-order function, both of which cooperatively reduce the peak demand for computing power.

[0019] In an embodiment of the present application, the resource control signal sent to the graphics processor driver layer in step S4 includes a virtual video memory paging strategy reset instruction, when the video memory allocation proportion 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, and at the same time, the on-demand return mechanism is activated to generate a video memory page replacement pipeline control signal, so that the video memory access delay characteristic signal is maintained within an acceptable range.

[0020] The present application also includes a digital twin platform GPU rendering resource dynamic scheduling system, comprising:

[0021] The acquisition module acquires the rendering task feature signal of the current frame in real time, and simultaneously acquires the GPU multi-dimensional resource consumption signal;

[0022] The coordination module generates the GPU resource demand prediction signal and the 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;

[0023] The comparison module generates a dynamic resource allocation instruction signal through a multi-objective optimization function according to a GPU resource demand prediction signal, a task dependency relationship signal, a user priority signal and a real-time resource bottleneck type signal.

[0024] The preview module executes the dynamic resource allocation instruction signal, sends a 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 meets the task dependency relationship signal.

[0025] The analysis module feeds 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, and generates a parameter self-optimization signal for updating the prediction and scheduling logic of the subsequent frame.

[0026] The present application provides a kind of digital twin platform GPU rendering resource dynamic scheduling method and system, by real-time acquisition rendering task characteristic signal and GPU multidimensional resource consumption signal, generates load prediction signal;Based on multi-objective optimization function fusion task dependency relationship, user priority and resource bottleneck signal, dynamically generates computing power quota allocation, video memory allocation proportion and drawing call order instruction;When executing instruction, it is forced to guarantee the timing of task dependency chain, and the actual scheduling effect signal is fed back to the prediction and optimization process, forms closed loop self-optimization mechanism, realizes that resource allocation is automatically adjusted with digital twin scene dynamic change, simultaneously meet the real-time demand of high priority task and fairness constraint between multiple users. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used for the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0028] Figure 1 It is a method flow chart of a kind of digital twin platform GPU rendering resource dynamic scheduling method;

[0029] Figure 2 It is a system architecture diagram of a kind of digital twin platform GPU rendering resource dynamic scheduling system. DETAILED DESCRIPTION

[0030] Following make the embodiments of the present application more clear by specific examples, and other advantages and effects of the present application can be easily understood by those skilled in the art from the disclosure of the present application. The present application can also be implemented or applied by other different embodiments, and various modifications or changes can be made based on different views and applications without departing from the spirit of the present application. It should be noted that the following examples and features in the examples can be combined with each other without conflict.

[0031] It should be noted that the diagrams provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component may be a random change, and the component layout pattern may be more complex.

[0032] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application, however, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details, and in other embodiments, the known structures and devices are shown in the form of block diagrams rather than in the form of details, to avoid making the embodiments of the present application difficult to understand.

[0033] Please refer to Figures 1-2 , which shows a digital twin platform GPU rendering resource dynamic scheduling method and system of the present application. The digital twin platform GPU rendering resource dynamic scheduling method of the present application includes: S1: real-time acquisition of rendering task characteristic signals of the current frame, and acquisition of GPU multi-dimensional resource consumption signals at the same time, 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 characteristic signals, the historical resource consumption signals and the scene dynamic change signals, generating the GPU resource demand prediction signal and the load fluctuation trend signal of the future frame; S3: according to the GPU resource demand prediction signal, the task dependency relationship signal, the user priority signal and the 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 order signal; S4: executing the dynamic resource allocation instruction signal, sending the 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 conforms to the task dependency relationship 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 the subsequent frame.

[0034] As Figure 1As shown, the core process of a digital twin platform GPU rendering resource dynamic scheduling method, the technical scheme forms a closed-loop control system through five steps. Step S1 first establishes a signal acquisition basis: real-time capture of the rendering task characteristic signal of the current frame, which is essentially a digital abstraction of the scene rendering load, specifically covering the number of geometric bodies, lighting calculation complexity, special effect particle density, and other implicit features that affect GPU load, while synchronously collecting GPU multi-dimensional resource consumption signals, which are characterized by simultaneous monitoring of computing power resource consumption signals (such as shader core occupancy), video memory resource consumption signals (such as texture buffer usage peak), and bandwidth resource consumption signals (such as video memory bus transmission volume). These signals collectively form the original input matrix for resource scheduling. Step S2 implements dynamic prediction: based on the real-time signals of step S1, combined with historical resource consumption signals (such as the computing power allocation records of the previous N frames) and scene dynamic change signals (such as view displacement vector, physical 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, where the load fluctuation trend signal specifically quantifies the resource demand change rate caused by scene mutations (such as the sharp increase in geometry data loading rate when the view angle changes rapidly). Step S3 completes intelligent decision-making: input the prediction signals output by step S2 and the task dependency signals of 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 video memory bandwidth utilization exceeding the threshold) into a multi-objective optimization function, which is mathematically based on Pareto frontier search, and outputs dynamic resource allocation instruction signals - specifically including computing power quota allocation signals (defining the proportion of computing units that each task can occupy), video memory allocation proportion signals (controlling the allocation weight of texture buffer and frame buffer), and rendering call order signals (arranging the task execution sequence to avoid dependency conflicts). Step S4 executes hardware control: convert the instruction signals generated in step S3 into resource control signals recognizable by the GPU driver layer, redirect the rendering call command stream through the graphics application programming interface, and strictly follow the task dependency signal to build a barrier synchronization mechanism (such as inserting a synchronization barrier command) during execution to ensure that the execution order of the dependent task chain meets the data consistency requirements of the rendering pipeline. Step S5 builds a feedback loop: collect the actual resource consumption signals (such as the actual usage time of computing power) and scheduling effect evaluation signals (such as the standard deviation of frame delay) of the current frame, and feed them back to the prediction process of step S2 and the optimization function of step S3 to generate parameter self-optimization signals (such as prediction model weight correction coefficients and optimization function constraint boundary adjustment amounts), driving the adaptive evolution of subsequent frame scheduling logic. The essence of this claim is to build a complete closed loop of "perception-prediction-decision-execution-feedback", and to achieve real-time synchronization of resource scheduling and digital twin scene changes through dynamic coupling of signal flow.

[0035] Further, the rendering task feature signal of step S1 is further expanded in depth, and two types of key signals are newly added to improve the load perception accuracy. The spatial position distribution feature signal of the scene object is generated by spatial grid subdivision technology: the visible space is divided into a three-dimensional grid, the bounding box density and the number of instances of the geometric body in each grid are counted, and a spatial aggregation heat map signal is generated. The signal is directly related to the locality of memory access - high-density areas need to be allocated bandwidth resources preferentially; the material texture complexity feature signal is generated by running time shader code analysis: the number of texture sampling instructions and the sampler level depth in the fragment shader are analyzed, the texture size and compression format (such as BC7 compression ratio) are combined to calculate the texture bandwidth demand coefficient, and the loop and branch complexity in the lighting calculation function are detected to generate a shader calculation intensity index. The two are fused to form the material texture complexity feature signal. The GPU multi-dimensional resource consumption signal further supplements the parallel utilization rate fluctuation signal of the graphics processor core: the occupancy rate change curve of the SIMD unit on the time axis is collected through the performance counter, and the periodic fluctuation characteristics of the algorithm resource utilization are identified; the memory access delay feature signal is generated by monitoring the memory controller queue depth and cache hit rate, and the memory subsystem response delay is quantified. These newly added signals constitute the space-time dynamic fingerprint of the rendering load: the spatial position distribution feature signal reveals the spatial correlation of the geometric data access mode, the material texture complexity feature signal decodes the calculation and storage overhead of surface shading, the parallel utilization rate fluctuation signal exposes the temporal imbalance of the algorithm resource, and the memory access delay feature signal reflects the storage level bottleneck position. In particular, the generation algorithm of the space-time dynamic fingerprint has the lightweight characteristic - the spatial grid density statistics adopts the adaptive subdivision of octree to avoid full scene traversal, and the shader analysis realizes zero runtime overhead through the pre-compiled intermediate representation instruction counting. The technical value of the claim lies in converting the ambiguous "scene complexity" into quantifiable physical indicators, so that the subsequent prediction and scheduling have fine-grained data support.

[0036] As Figure 1As shown, the multi-objective optimization function design of step S3 realizes the intelligent trade-off of the scheduling strategy by defining three conflicting optimization objectives. The first objective is to minimize the frame rendering delay jitter signal: this signal is generated by calculating the standard deviation of the GPU execution time of historical frames, specifically collecting the time interval from the draw call submission to the frame buffer exchange of consecutive M frames, calculating the coefficient of variation as the jitter quantization value, and the optimization function constrains the upper limit of this signal to ensure smooth rendering. The second objective is to maximize the high-priority task resource guarantee rate signal: for critical tasks marked by the user priority signal (such as security monitoring views), calculate the ratio of actual obtained computing power to applied computing power (e.g. if 45% of the actual obtained computing power is applied for 50%, the guarantee rate is 90%), and generate a comprehensive guarantee rate index combined with the memory allocation satisfaction degree (actual allocated memory / applied memory), and the optimization function requires that this signal is not less than a preset threshold. The third objective is to balance the resource fairness signal among multiple users: use the Jain fairness index algorithm, input the memory allocation proportion and computing power quota obtained by each user group, output the resource allocation difference coefficient (0 for complete fairness, 1 for extreme unfairness), and the optimization function optimizes towards the direction of the difference coefficient approaching 0. The solving mechanism of the multi-objective optimization function uses the 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 transformed into a minimum satisfaction rate constraint (e.g. ≥85%), and the resource fairness signal is used as the main optimization target for minimization search. The generation process of the dynamic resource allocation instruction signal contains three stages: first, select the optimization strategy according to the real-time resource bottleneck type signal - focus on delay jitter control when computing power bottleneck, and strengthen fairness guarantee when memory bandwidth bottleneck; second, generate a candidate instruction set based on the Pareto optimal solution space; finally, weight and sort the candidate solutions according to the user priority signal (e.g. monitoring view weight = 0.7, normal view weight = 0.3), and select the instruction with the highest total score as the output. The technical breakthrough of this claim lies in solving the "three difficult dilemmas" in resource allocation - traditional methods cannot simultaneously satisfy the requirements of low delay, high priority guarantee, and multi-user fairness, while this scheme realizes dynamic balance through multi-objective collaborative optimization.

[0037] Further, step S4 performs a key optimization mechanism in the process, which solves the problem of task dependency guarantee and resource bottleneck adaptation. When executing the rendering call sequence signal, first, a directed acyclic graph topology sequence signal is constructed according to the task dependency relationship signal, the rendering task is abstracted as a node, and the data flow relationship is expressed through a directed edge. For example, the vertex processing task node must be located before the fragment processing task node to form an execution sequence constraint, so as to ensure that the output data of the previous stage is complete and ready before starting the subsequent task. When a GPU write dependency scenario is detected, that is, the rendering result output by a task will be used as the input texture of another task, a GPU barrier synchronization signal is automatically inserted between the tasks. The signal is converted into a memory isolation command recognizable by the underlying hardware to force the GPU write queue to be refreshed, eliminating the risk of data race caused by out-of-order execution. More breakthrough is the intelligent batch processing mechanism of resource bottleneck perception: when the real-time resource bottleneck type signal continuously reports that the GPU bandwidth utilization rate exceeds the warning threshold, spatial clustering analysis is immediately started - the spatial proximity of the bounding boxes of the geometric objects in the scene is calculated, and the object group that meets the distance threshold is generated aggregation instruction, and the independent rendering call command is combined to form a batch optimization instruction signal. This signal triggers three optimization operations: the scattered vertex data is reorganized into a continuous storage vertex buffer and a data rearrangement signal is generated; textures with similar material properties are packed into a texture array and a texture mapping signal is generated; and multiple independent rendering calls are replaced by a single instantiation rendering and an instance expansion signal is generated. This mechanism significantly reduces the frequency of graphics application programming interface calls and the pressure on the GPU bus. In actual measurement, the number of interface calls is reduced by more than 70 percent, and the data transmission volume is reduced by nearly half. The execution process synchronously monitors the integrity of the dependency chain, and if the batch processing breaks the original dependency relationship, the batch processing group is dynamically split and a compensatory synchronization point is injected. The final output resource control signal contains a four-dimensional instruction set: sequence index instruction defines the task execution flow sequence, memory barrier coordinate instruction marks the GPU synchronization position, batch processing parameter instruction carries the instantiation scale and texture offset, and exception rollback instruction ensures that the original state is quickly restored when the dependency is broken. The essence of the innovation of this claim lies in realizing the autonomous trade-off of the execution layer: maximizing batch processing benefits when resources are scarce, while ensuring rendering correctness through dynamic dependency maintenance.

[0038] As Figure 1As shown, the feedback precision of step S5 constructs a quantitative evaluation system to drive system self-optimization. The scheduling effect evaluation signal contains two types of core indicators: the resource allocation imbalance degree 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, calculates the weighted mean of the computing power allocation deviation rate and the video memory allocation deviation rate, and forms a global resource imbalance index. The prediction deviation signal adopts a multi-error fusion strategy: the sliding window statistical error is calculated for the computing power demand prediction value sequence and the actual consumption value sequence, and the video memory bandwidth demand prediction adopts the percentage absolute error algorithm, and finally the prediction deviation index is generated by adaptive weight synthesis. The generation of the parameter self-optimization signal follows the double-path regulation principle: the first path acts on the prediction model, and the feature weight coefficient is dynamically adjusted according to the prediction deviation signal. When a specific feature signal such as material texture complexity continuously presents prediction deviation, the decision weight in the model is increased. When a dynamic signal such as view angle switching acceleration error fluctuates sharply, noise interference is suppressed by applying smoothing attenuation. The second path acts on the optimization function, and the resource allocation imbalance degree signal is used to reconstruct the constraint boundary: if the actual resource guarantee rate of high-priority tasks exceeds the tolerance interval compared with the planned value, the guarantee rate constraint lower limit is tightened; if the frame rendering delay jitter breaks through the threshold, an additional jitter suppression penalty factor is injected. The self-optimization process introduces a stability protection mechanism: only when the evaluation signal of consecutive multiple frames exceeds the standard can the parameter update be triggered, avoiding oscillation caused by instantaneous fluctuations; the parameter adjustment amplitude uses a composite calculation model of historical deviation accumulation and current change trend to ensure that the correction step is both responsive and smooth and controllable. The technical breakthrough of this claim lies in converting abstract system efficiency into a quantifiable signal system, enabling the closed-loop system to have continuous evolution capability.

[0039] Specifically, the prediction signal generation quality of step S2 improves the anti-interference ability through signal preprocessing. The historical resource consumption signal is normalized by a sliding window: a time window containing a certain number of frames is set, and each resource consumption data in the window is standardized based on statistical characteristics to eliminate the difference in data magnitude caused by scene initialization or sudden events, and a denoised resource consumption baseline signal is generated. The scene dynamic change signal uses 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 the persistent state transition such as gentle view movement, and the high-frequency component corresponds to the sudden event such as view rapid turning or physical explosion. The prediction model only takes the high-frequency disturbance component signal as the input because it has a strong correlation with the load mutation; the low-frequency trend component signal is directly superimposed on the prediction output end to generate a steady-state resource demand signal. In specific implementation, the high-frequency disturbance component drives the short-time prediction branch: a dynamic coefficient moving average model is used to update the load fluctuation trend signal of future frames, and the response speed of the model is positively correlated with the disturbance intensity - the more intense the disturbance, the faster the model adjusts. The low-frequency trend component drives the long-time prediction branch: a linear regression fitting is used to generate a steady-state resource demand baseline. The final output load fluctuation trend signal is composed of the high-frequency prediction result and the low-frequency baseline signal. The core value of this claim lies in significantly improving the robustness of the prediction system: the normalization process eliminates the scale difference of historical data, and the frequency domain decomposition isolates the characteristics of sudden events, so that the prediction signal remains stable and accurate in the complex environment of digital twin.

[0040] As shown in Figure 2 The acquisition module acquires the rendering task feature signal of the current frame in real time, and simultaneously acquires the GPU multi-dimensional resource consumption signal; the coordination module generates the GPU resource demand prediction signal and the 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 a multi-objective optimization function according to the GPU resource demand prediction signal, the task dependency relationship 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 meets the task dependency relationship signal; the analysis module feeds 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, and generates a parameter self-optimization signal for updating the prediction and scheduling logic of the subsequent frame.

[0041] As shown in Figure 2As shown, the real-time resource bottleneck signal of step S3 designs a dynamic degradation strategy to solve the emergency processing problem of sudden shortage of computing power resources. When the real-time resource bottleneck type signal continuously indicates that the computing power resources are limited, the dynamic resource allocation instruction signal activates the progressive detail level regulation mechanism. The mechanism includes a double-path cooperative degradation: a geometric simplification path generates a geometric simplification instruction signal to drive a grid simplification algorithm to generate a multi-level detail version of a high polygon model, and dynamically select the detail level according to the user priority signal - the key user task maintains the original detail, and the simplified model is used for the ordinary user task, and the simplification strength is proportional to the computing power gap. The shader optimization path generates a shader complexity regulation signal, which realizes through a runtime shader replacement technology: pre-compile different complexity shader variants, detect high-overhead computing functions such as real-time soft shadows or ambient light occlusion, and replace them with approximate low-order functions such as pre-computed light maps, and the replacement granularity is dynamically controlled by the user priority signal and the real-time computing power pressure. The double-path degradation depth linkage: geometric simplification reduces vertex processing load, and shader optimization reduces fragment calculation pressure, and the two cooperate to suppress the peak value of computing power demand within the range of available resources. The degradation process implements visual fidelity protection: the high detail is maintained in the center area of the user's viewport, and stronger simplification is implemented in the edge area; the geometric detail of the object with high motion speed is reduced, but the texture accuracy is maintained. When the computing power resource recovery signal arrives, a progressive backtracking strategy is adopted to gradually improve the detail level frame by frame to avoid visual mutation. The innovation of the claim lies in establishing an elastic degradation system, which resolves the computing power crisis by multi-dimensional resource reduction under the premise of guaranteeing the real-time performance of key tasks.

[0042] The digital twin platform GPU rendering resource dynamic scheduling method and system of the present application can solve the problem of low GPU resource utilization and difficulty in balancing real-time performance under dynamic load.

[0043] Therefore, by the digital twin platform GPU rendering resource dynamic scheduling method and system of the present application, the problem of low GPU resource utilization and difficulty in balancing real-time performance under dynamic load can be solved.

[0044] The above embodiments are only illustrative of the principles of the present application and its efficacy, and are not intended to limit the present application. Any modification or change made by any person skilled in the art without departing from the spirit and scope of the present application shall be covered by the claims of the present application.

Claims

1. A method for dynamically scheduling GPU rendering resources of a digital twin platform, characterized in that, Comprise: S1: real-time acquisition of the rendering task feature signal of the current frame, while collecting GPU multi-dimensional resource consumption signal, the multi-dimensional resource includes computing power resource consumption signal, video memory resource consumption signal and bandwidth resource consumption signal; S2: based on the rendering task feature signal, historical resource consumption signal and scene dynamic change signal, through the prediction model feature weight coefficient, the GPU resource demand prediction signal and the load fluctuation trend signal of the future frame are generated; In step S2, when the load fluctuation trend signal is generated, the scene dynamic change signal is introduced, including the view switching acceleration signal and the physical simulation real-time update signal, wherein the view switching acceleration signal is generated by the differential operation of the continuous frame view matrix transformation, and the physical simulation real-time update signal is generated by detecting the rigid body motion trajectory deviation and the particle system emission rate change, 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 rendering load step mutation caused by sudden interactive events; S3: 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 dynamic resource allocation instruction signal is generated through the multi-objective optimization function, which includes the computing power quota allocation signal, the video memory allocation proportion signal and the drawing call order signal; 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 between multiple users as the core optimization target, wherein the frame rendering delay jitter signal is generated by calculating the standard deviation of the historical frame completion time, the high priority task resource guarantee rate signal is generated by calculating the ratio of the actual computing power obtained by the key rendering task to the applied computing power, and the resource fairness signal is generated by evaluating the video 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 target; S4: execute the dynamic resource allocation instruction signal, send 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 conforms to the task dependency signal; S5: the actual resource consumption signal of the current frame and the scheduling effect evaluation signal are fed back to the resource demand prediction and dynamic resource allocation process, and the parameter self-optimization signal is generated for updating the prediction and scheduling logic of the subsequent frame.

2. The method of claim 1, wherein, The rendering task feature signal collected in step S1 further includes the spatial position distribution feature signal of the scene object and the material texture complexity feature signal, wherein the spatial position distribution feature signal is generated by analyzing the aggregation density of the scene object in the visible space, and the material texture complexity feature signal is generated by analyzing the instruction complexity and texture sampling frequency of the shader program, and the GPU multi-dimensional resource consumption signal collected at the same time also includes the parallel utilization rate fluctuation signal of the graphics processor core and the video memory access delay feature signal, which together constitute the space-time dynamic fingerprint of the rendering load.

3. The method of claim 1, wherein, The step S4 executes the drawing call sequence signal, constructs a directed acyclic graph topology sequence signal according to the task dependency relationship signal, forcibly requires the vertex shading task output signal to be executed before the fragment shading task input signal, inserts a memory barrier synchronization signal for the task with a memory write dependency relationship, and automatically combines spatially adjacent drawing call commands to generate a batch optimization instruction signal to reduce the number of graphic application program interface calls when the real-time resource bottleneck type signal indicates that the memory bandwidth is limited.

4. The GPU rendering resource dynamic scheduling method of a digital twin platform according to claim 1, characterized in that, The scheduling effect evaluation signal in the step S5 includes a resource allocation imbalance degree signal and a prediction deviation degree signal, wherein the resource allocation imbalance degree signal is generated by comparing the absolute value of the residual of the computing power quota allocation signal and the actual consumption signal, the prediction deviation degree signal is generated by calculating the root mean square error of 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 the step S2 and the multi-objective optimization function constraint boundary in the step S3 according to the prediction deviation degree signal.

5. The GPU rendering resource dynamic scheduling method of a digital twin platform according to claim 1, characterized in that, Before generating the resource demand prediction signal in the 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 sudden events, and the low-frequency trend component signal is directly superimposed on the prediction result to generate a steady-state resource demand signal.

6. The GPU rendering resource dynamic scheduling method of a digital twin platform according to claim 1, characterized in that, When the real-time resource bottleneck type signal indicates that the computing power resource is limited, the dynamic resource allocation instruction signal starts a gradual detail level degradation strategy based on the user priority signal, generates a geometry simplification instruction signal and a shader complexity control signal, wherein the geometry simplification instruction signal drives a grid simplification algorithm to dynamically generate a low-polygon version model, and the shader complexity control signal is realized by replacing a high-overhead lighting calculation function with an approximate low-order function, both of which cooperatively reduce the peak value of the computing power demand.

7. The GPU rendering resource dynamic scheduling method of a digital twin platform according to claim 1, characterized in that, The resource control signal sent by the step S4 to the graphic processor driver layer includes a virtual memory paging strategy reset instruction, when the memory allocation proportion 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, and a demand-driven return mechanism is activated to generate a memory page replacement pipeline control signal, so that the memory access delay characteristic signal is maintained within an acceptable range.

8. A rendering resource dynamic scheduling system using the rendering resource dynamic scheduling method of the digital twin platform GPU rendering resource of any one of claims 1-7, characterized in that, Comprise: The acquisition module acquires the rendering task feature signal of the current frame in real time, and simultaneously acquires the GPU multi-dimensional resource consumption signal; The coordination module generates the GPU resource demand prediction signal and the 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 a multi-objective optimization function according to the GPU resource demand prediction signal, the task dependency relationship signal, the user priority signal and the real-time resource bottleneck type signal; a rehearsal module, which executes the dynamic resource allocation instruction signal, sends a resource control signal to a GPU driver layer to adjust rendering task execution logic, and ensures that the execution order of a dependent task chain conforms to the task dependency relationship signal; an analysis module, which feeds actual resource consumption signal of a current frame and scheduling effect evaluation signal to a resource demand prediction and dynamic resource allocation process, and generates a parameter self-optimization signal for updating prediction and scheduling logic of a subsequent frame.

Images