A task model training method for computing power scheduling

Abstract

This invention discloses a task model training method for computing power scheduling, comprising the following steps: collecting and preprocessing data to generate standardized task sequences and standardized computing power state sequences; generating task feature sequences based on the standardized task sequences and computing power feature sequences based on the standardized computing power state sequences; performing simultaneous time-step concatenation to generate a task-computing power joint state sequence; inputting the task-computing power joint state sequence into an improved S4 state-space sequence model to generate a predicted computing power state evolution sequence; performing multi-step evolutionary consistency constraint training on the predicted computing power state evolution sequence to generate a consistency loss value; updating model parameters and outputting the task-computing power scheduling mapping result. This invention achieves stable training of the computing power scheduling model and multi-step evolutionary consistency prediction, improving scheduling stability and computing power resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of computing power resource scheduling technology, and in particular to a task model training method for computing power scheduling. Background Technology

[0002] With the widespread deployment of cloud computing, edge computing, and heterogeneous computing architectures, the frequency and complexity of scheduling computing tasks among various types of computing resources continue to increase. Existing computing scheduling technologies mostly rely on modeling information such as task size and resource load, and generate scheduling results through rule-based strategies, heuristic algorithms, or learning-based models to meet the basic requirements of task execution efficiency and resource utilization. Among related technologies, some solutions introduce sequence modeling methods to predict task or computing power status, while others combine supervised learning or reinforcement learning mechanisms to optimize scheduling strategies to adapt to dynamically changing operating environments.

[0003] Existing computing power scheduling techniques based on learning models typically use task characteristics and computing power status as static inputs, directly outputting scheduling decisions or scheduling scores. The model training process focuses on single-step prediction errors or immediate scheduling benefits, paying insufficient attention to the evolution of computing power status over continuous time steps. These methods do not impose constraints on the multi-step evolution relationship between predicted and actual computing power status during the training phase, leading to accumulated state shifts in the model during long-term scheduling and insufficient stability of scheduling results. Although some sequence modeling schemes introduce state-space model structures, the state transition parameters are fixed, making it difficult to reflect the different impacts of different task conditions on the evolution path of computing power status.

[0004] Meanwhile, existing technologies handle the time scale of computing power state updates in a relatively simple way. Task load mutations and computing power resource capacity evolution are often modeled through a unified state update mechanism, lacking the distinction between fast and slow timescale state changes. During model training, there is a lack of a unified constraint mechanism between the scheduling target supervision signal and the state prediction error. Model parameter updates rely more on single target optimization, making it difficult to balance the consistency of computing power state evolution and the achievement of scheduling targets. These problems limit the applicability and generalization ability of existing computing power scheduling models in complex, multi-step scheduling scenarios.

[0005] Therefore, how to provide a task model training method for computing power scheduling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a task model training method for computing power scheduling. This invention introduces structured state space modeling and a multi-step evolution consistency constraint mechanism to jointly model and train the temporal evolution process of task characteristics and computing power state, thereby realizing computing power state prediction and scheduling mapping generation for multiple time scales. It has the advantages of high scheduling stability, strong evolution consistency and excellent generalization ability.

[0007] A task model training method for computing power scheduling according to an embodiment of the present invention includes the following steps: Collect data on tasks to be scheduled and computing resource operation data, and perform preprocessing on the data to be scheduled and computing resource operation data to generate standardized task sequences and standardized computing power status sequences; Based on the standardized task sequence, task computation load characteristics and task latency constraint characteristics are extracted to generate a task feature sequence. Based on the standardized computing power status sequence, computing power load characteristics and computing power availability characteristics are extracted to generate a computing power feature sequence. Perform time-synchronous splicing and association annotation on the task feature sequence and the computing power feature sequence to generate a joint task-computing power state sequence; The task-computing power joint state sequence is input into the improved S4 state space sequence model, which includes three parts: task condition state matrix generation, dual time-scaled state update gating, and step residual state advancement, and outputs the predicted computing power state evolution sequence. A windowed reference computing power state evolution sequence is constructed based on the standardized computing power state sequence. Multi-step evolution consistency constraint training is performed on the predicted computing power state evolution sequence and the reference computing power state evolution sequence to generate a consistency loss value. Based on the consistency loss value and the scheduling target supervision signal, the parameters of the improved S4 state space sequence model are iteratively updated to obtain the convergent model parameter set. Based on the convergent model parameter set, the task-computing power joint state sequence is inferred and calculated, and the task-computing power scheduling mapping result is output.

[0008] Optionally, the task data to be scheduled includes task submission timestamps, task computation scale parameters, task parallelism parameters, and task latency constraint parameters, and an original task sequence is constructed according to the task submission order; the computing power resource operation data includes computing load parameters, available computing capacity parameters, and resource utilization parameters under the corresponding timestamp of the computing power node, and an original computing power state sequence is constructed according to the time order; the preprocessing includes performing timestamp alignment processing on the original task sequence and the original computing power state sequence to form a time-consistent data sequence; performing missing value imputation and outlier removal processing on the time-aligned data sequence; performing dimensional normalization processing on the processed original task sequence to generate a standardized task sequence; and performing dimensional normalization processing on the processed original computing power state sequence to generate a standardized computing power state sequence.

[0009] Optionally, the process of generating the task feature sequence and computing power feature sequence includes: For each task time step in the standardized task sequence, the task computational load intensity value is calculated based on the task computational scale parameter and the task parallelism parameter. The task computational load intensity values ​​are arranged in chronological order to form a task computational load characteristic sequence. For each task time step in the standardized task sequence, the remaining schedulable time window length of the task is calculated based on the task delay constraint parameters and the corresponding task submission timestamp. The remaining schedulable time window lengths of the tasks are arranged in chronological order to form a task delay constraint feature sequence. The task computation load feature sequence and the task latency constraint feature sequence are concatenated along the feature dimension to generate the task feature sequence; For each time step in the standardized computing power state sequence, the computing power load occupancy ratio is calculated based on the computing power node's calculated load parameters and available computing capacity parameters, and arranged in chronological order to form a computing power load characteristic sequence. For each time step in the standardized computing power status sequence, the computing power availability index value is calculated based on the available computing capacity parameter and the resource utilization rate parameter, and arranged in chronological order to form a computing power availability characteristic sequence. The computing load feature sequence and the computing availability feature sequence are concatenated along the feature dimension to generate the computing power feature sequence.

[0010] Optionally, the process of generating the task-computing power joint state sequence includes: On a unified time axis, determine the corresponding time step index of the task feature sequence and the computing power feature sequence, and match the task feature vector and computing power feature vector under the same time step index one by one. The matching task feature vector and computing power feature vector are concatenated along the feature dimension to generate a joint feature vector for the corresponding time step. Add a time step identifier and a task identifier to each joint feature vector to form a joint state term that includes time sequence information and task association information; Arrange the joint state items of each time step in the order of time step index to generate the task-computing power joint state sequence.

[0011] Optionally, the task condition state matrix generation step in the improved S4 state-space sequence model includes: At each time step, the task feature vector and computing power feature vector of the corresponding time step are read from the task-computing power joint state sequence, and the task feature vector is used as the conditional input vector. Based on the conditional input vector, perform linear mapping and nonlinear transformation to generate task conditional embedding vectors; Based on the task condition embedding vector, perform element-wise modulation operation on the preset basic state matrix parameters to generate a task condition state matrix corresponding to the current task features; Structural constraints are imposed on the task condition state matrix, maintaining a fixed order in the state dimension but allowing it to change with the task conditions in the time step dimension. The generated task condition state matrix is ​​then used as the state transition matrix for the current time step.

[0012] Optionally, the dual-time-scaled state update gating step in the improved S4 state-space sequence model includes: At each time step, based on the task-computing power joint state vector of the current time step, a fast time scale update control vector and a slow time scale update control vector are generated respectively. Based on the fast timescale update control vector, high-frequency update calculation is performed on the state increment under the action of the task condition state matrix corresponding to the current time step to obtain the fast timescale state update component, which characterizes the transient computing power state response caused by changes in task load. Based on the slow time-scaled update control vector, a low-frequency smooth update calculation is performed on the accumulated state representation of the previous time step to obtain the slow time-scaled state update component, which characterizes the gradual evolution process of computing power resource capacity and occupancy level. Gating coefficient constraints are applied to the fast time-scaled state update component and the slow time-scaled state update component respectively, so that the two types of update components participate in the state update in the same time step according to a preset ratio; The fast timescale state update component and the slow timescale state update component after gating constraints are weighted and fused to generate the comprehensive state update result for the current time step. The comprehensive state update result is used as the input state for the step residual state advancement.

[0013] Optionally, the step-by-step residual state advancement step in the improved S4 state-space sequence model includes: At each time step, the candidate state representation for the current time step is generated based on the comprehensive state update result of the dual-time-scaled state update gate output. Read the historical state representation from the task-computing power joint state sequence, which is at least one time step away from the current time step, and use the historical state representation as the step residual term; Apply time step interval weight decay to the step residual term; The cross-step residual term after weight decay is added to the candidate state representation of the current time step element by element to generate the cross-step residual fused state representation; Using the step residual fusion state representation as the initial state input for the next time step, the process is continuously advanced along the time step direction to form a predictive computing power state evolution sequence.

[0014] Optionally, the step of performing multi-step evolutionary consistency constraint training and generating consistency loss values ​​includes: In each training iteration, the prediction step interval corresponding to the windowed reference computing power state evolution sequence is determined, and multi-step continuous state segments of the prediction computing power state evolution sequence are extracted within the prediction step interval. The multi-step continuous state segments are aligned one by one with the reference computing power state evolution sequence within the corresponding time window according to the time step order to form a multi-step state pairing sequence. For each pair of time-step aligned predicted computing power states and reference computing power states, perform state-by-state dimension difference calculation to obtain the state deviation vector for the corresponding time step; perform norm calculation on the state deviation vector in the state dimension to generate the state deviation scalar for the corresponding time step. The state deviation scalars of each time step are cumulatively weighted according to the time step sequence, with higher weights applied to the state deviation scalars corresponding to more distant prediction time steps, to obtain the cumulative value of multi-step evolution consistency. The cumulative value of multi-step evolution consistency is determined as the consistency loss value, and the consistency loss value is input into the model parameter update process to constrain the evolution consistency of the prediction computing power state evolution sequence on multi-step time scales.

[0015] Optionally, the step of iteratively updating the parameters of the improved S4 state-space sequence model based on the consistency loss value and the scheduling target supervision signal to obtain the convergent model parameter set includes: In each training iteration, the consistency loss value obtained from the multi-step evolution consistency constraint training and the scheduling supervision loss value calculated from the scheduling target supervision signal are obtained respectively. The consistency loss value and the scheduling supervision loss value are weighted and fused according to a preset weight ratio to generate a joint training loss value. Based on the joint training loss value, reverse update calculations are performed on the parameters related to the generation of the task conditional state matrix, the parameters related to the gating of the dual time-scaled state update, and the parameters related to the propagation of the step residual state in the improved S4 state-space sequence model. The parameter update step is repeated in several consecutive training iterations, and the corresponding joint training loss value is recorded after each iteration. When the change in the joint training loss value is lower than a preset threshold during a preset number of consecutive training iterations, the model parameters are determined to have reached a convergent state, parameter updates are stopped, and the model parameters obtained at the time of stopping updates are determined as the convergent model parameter set.

[0016] The beneficial effects of this invention are: (1) By introducing a task model training method based on structured state space, the joint state sequence of task features and computing power state is time-series modeled, and consistency constraints are applied to the multi-step evolution process of computing power state during training, so that the model can maintain consistency between the computing power state prediction and the actual evolution trend in continuous time steps, avoid the accumulation of state error over time during scheduling, and improve the stability and reliability of computing power scheduling results in long-term running scenarios.

[0017] (2) By constructing a task condition state matrix generation mechanism and a dual time-scale state update gating mechanism in the state space sequence model, the computing power state update process can distinguish between the rapid changes caused by task load and the slow changes caused by the evolution of computing power resource capacity, thereby realizing fine modeling of state changes at different time scales and improving the model's adaptability to complex task load changes and heterogeneous computing power resource operation characteristics.

[0018] (3) By jointly introducing multi-step evolution consistency loss and scheduling target supervision signal during the model training stage, the parameters related to task condition state matrix generation, dual time-scale state update gating and step residual state advancement are optimized in a coordinated manner, so that the model can take into account scheduling target constraints while ensuring the consistency of computing power state evolution, thereby improving the generalization performance and scheduling rationality of task-computing power scheduling mapping results. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a task model training method for computing power scheduling proposed in this invention; Figure 2 This is a schematic diagram illustrating the process of constructing standardized task sequences and standardized computing power state sequences in a task model training method for computing power scheduling proposed in this invention. Figure 3 This is a schematic diagram of the process for generating a joint state sequence of task and computing power in a task model training method for computing power scheduling proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for training a task model for computing power scheduling includes the following steps: Collect data on tasks to be scheduled and computing resource operation data, and perform preprocessing on the data to be scheduled and computing resource operation data to generate standardized task sequences and standardized computing power status sequences; Based on the standardized task sequence, task computation load characteristics and task latency constraint characteristics are extracted to generate a task feature sequence. Based on the standardized computing power status sequence, computing power load characteristics and computing power availability characteristics are extracted to generate a computing power feature sequence. Perform time-synchronous splicing and association annotation on the task feature sequence and the computing power feature sequence to generate a joint task-computing power state sequence; The task-computing power joint state sequence is input into the improved S4 state space sequence model, which includes three parts: task condition state matrix generation, dual time-scaled state update gating, and step residual state advancement, and outputs the predicted computing power state evolution sequence. A windowed reference computing power state evolution sequence is constructed based on the standardized computing power state sequence. Multi-step evolution consistency constraint training is performed on the predicted computing power state evolution sequence and the reference computing power state evolution sequence to generate a consistency loss value. The improved S4 state-space sequence model parameters are iteratively updated based on the consistency loss value and the scheduling target supervision signal to obtain a converged model parameter set. Inference calculations are then performed on the task-computing power joint state sequence based on this converged model parameter set, outputting the task-computing power scheduling mapping result. Specifically, after model training is completed, the converged model parameter set is fixed, and the task-computing power joint state sequence is input into the improved S4 state-space sequence model in time step order. Task conditional state matrix generation, dual-timescale state update gating, and cross-step residual state advancement operations are performed sequentially to obtain the predicted computing power state evolution result for the corresponding time step, thereby completing the inference output of the task-computing power scheduling mapping result.

[0022] In this embodiment, the task data to be scheduled includes task submission timestamps, task computation scale parameters, task parallelism parameters, and task latency constraint parameters, and an original task sequence is constructed according to the task submission order; the computing power resource operation data includes computing load parameters, available computing capacity parameters, and resource utilization parameters under the corresponding timestamp of the computing power node, and an original computing power state sequence is constructed according to the time order; the preprocessing includes performing timestamp alignment processing on the original task sequence and the original computing power state sequence to form a time-consistent data sequence; performing missing value imputation and outlier removal processing on the time-aligned data sequence; performing dimensional normalization processing on the processed original task sequence to generate a standardized task sequence; and performing dimensional normalization processing on the processed original computing power state sequence to generate a standardized computing power state sequence.

[0023] In this embodiment, the generation process of the task feature sequence and computing power feature sequence includes: For each task time step in the standardized task sequence, the task computational load intensity value is calculated based on the task computational scale parameter and the task parallelism parameter. The task computational load intensity values ​​are arranged in chronological order to form a task computational load characteristic sequence. The computational load intensity value is calculated by multiplying and dividing the task computational scale parameter and the task parallelism parameter for each task time step to convert it into the computational requirement per unit time, and the result is normalized to obtain the task computational load intensity value. For each task time step in the standardized task sequence, the remaining schedulable time window length is calculated based on the task delay constraint parameters and the corresponding task submission timestamp. The remaining schedulable time window lengths are then arranged in chronological order to form a task delay constraint feature sequence. Specifically, the calculation of the remaining schedulable time window length involves taking the task delay constraint parameters as the deadline for each task time step, calculating the difference between the deadline and the task submission timestamp, and subtracting the waiting time to obtain the remaining schedulable time window length. The task computation load feature sequence and the task latency constraint feature sequence are concatenated along the feature dimension to generate the task feature sequence; For each time step in the standardized computing power state sequence, the computing power load occupancy ratio is calculated based on the computing power node's computing load parameters and available computing capacity parameters, and arranged in chronological order to form a computing power load characteristic sequence. Specifically, the calculation of the computing power load occupancy ratio involves performing a ratio calculation on the computing power node's computing load parameters and available computing capacity parameters for each time step, and then truncating the ratio to ensure it falls within a preset range to obtain the computing power load occupancy ratio. For each time step in the standardized computing power status sequence, the computing power availability index value is calculated based on the available computing capacity parameter and the resource occupancy rate parameter, and arranged in chronological order to form a computing power availability feature sequence. Specifically, the computing power availability index value is calculated for each time step by taking the available computing capacity parameter as the effective supply and the resource occupancy rate parameter as the occupancy penalty term, and then weighting and normalizing the effective supply and the occupancy penalty term to obtain the computing power availability index value. The computing load feature sequence and the computing availability feature sequence are concatenated along the feature dimension to generate the computing power feature sequence.

[0024] In this embodiment, the process of generating the task-computing power joint state sequence includes: On a unified time axis, determine the corresponding time step index of the task feature sequence and the computing power feature sequence, and match the task feature vector and computing power feature vector under the same time step index one by one. The matching task feature vector and computing power feature vector are concatenated along the feature dimension to generate a joint feature vector for the corresponding time step. Add a time step identifier and a task identifier to each joint feature vector to form a joint state term that includes time sequence information and task association information; Arrange the joint state items of each time step in the order of time step index to generate the task-computing power joint state sequence.

[0025] In this embodiment, the task condition state matrix generation step in the improved S4 state-space sequence model includes: At each time step, the task feature vector and computing power feature vector of the corresponding time step are read from the task-computing power joint state sequence, and the task feature vector is used as the conditional input vector. Based on the conditional input vector, perform linear mapping and nonlinear transformation to generate task conditional embedding vectors; Based on the task condition embedding vector, perform element-wise modulation operation on the preset basic state matrix parameters to generate a task condition state matrix corresponding to the current task features; A structured constraint is imposed on the task condition state matrix, maintaining a fixed order in the state dimension while allowing it to change with task conditions in the time step dimension. The generated task condition state matrix is ​​then used as the state transition matrix for the current time step. Specifically, when generating the task condition state matrix, it is restricted to a block structure matrix composed of several predefined state sub-blocks arranged in a fixed topological relationship. The dimension of each state sub-block remains unchanged throughout the training process, and only the numerical parameters of each state sub-block are allowed to change continuously with the task condition embedding vector. This achieves adaptive modulation of the state transition relationship to different task conditions while ensuring that the order of the state space remains constant.

[0026] In this embodiment, the dual-time-scaled state update gating step in the improved S4 state-space sequence model includes: At each time step, based on the task-computing power joint state vector of the current time step, a fast time scale update control vector and a slow time scale update control vector are generated respectively. Based on the fast timescale update control vector, high-frequency update calculation is performed on the state increment under the action of the task condition state matrix corresponding to the current time step to obtain the fast timescale state update component, which characterizes the transient computing power state response caused by changes in task load. The high-frequency update calculation includes: within each time step, based on the task-computing power joint state vector and the task condition state matrix of the current time step, performing an instantaneous state increment calculation without time smoothing term on the state variables, so that the state update can respond quickly to sudden changes in task load. Based on the slow timescale update control vector, a low-frequency smooth update calculation is performed on the accumulated state representation of the previous time step to obtain the slow timescale state update component, which characterizes the gradual evolution process of computing power resource capacity and occupancy level. The low-frequency smooth update calculation includes: in each time step, based on the state representation of the previous time step and the task-computing power joint state vector of the current time step, a cumulative update calculation with a time smoothing factor is performed on the state variables to make the state change maintain a gradual transition between consecutive time steps. Gating coefficient constraints are applied to the fast-timescale state update component and the slow-timescale state update component respectively, so that the two types of update components participate in the state update in the same time step according to a preset ratio. The gating coefficient constraint includes: after obtaining the fast-timescale state update component and the slow-timescale state update component, assigning a gating coefficient with a limited numerical range to the two types of update components, and controlling the scaling of the update component amplitude so that the state update intensity in a single time step is within the preset constraint range. Participating in the state update in the preset ratio includes: weighting and synthesizing the fast-timescale state update component and the slow-timescale state update component according to the corresponding gating coefficient, so that the two types of update components jointly participate in the final state update of the current time step according to the ratio determined by the gating coefficient. The fast timescale state update component and the slow timescale state update component after gating constraints are weighted and fused to generate the comprehensive state update result for the current time step. The comprehensive state update result is used as the input state for the step residual state advancement.

[0027] In this embodiment, the step-by-step residual state advancement step in the improved S4 state-space sequence model includes: At each time step, the candidate state representation for the current time step is generated based on the comprehensive state update result of the dual-time-scaled state update gate output. Read the historical state representation from the task-computing power joint state sequence, which is at least one time step away from the current time step, and use the historical state representation as the step residual term; Apply time step interval weight decay processing to the step residual term; for each step residual term, according to the time step interval length between the historical state representation corresponding to the step residual term and the current time step, assign a weight coefficient that decreases monotonically with the time step interval, and apply the weight coefficient to the corresponding historical state representation so that the historical state with a larger time step interval has a lower influence on the advancement of the current state. The cross-step residual term after weight decay is added to the candidate state representation of the current time step element by element to generate the cross-step residual fused state representation; Using the step residual fusion state representation as the initial state input for the next time step, the process is continuously advanced along the time step direction to form a predictive computing power state evolution sequence.

[0028] In this embodiment, the step of performing multi-step evolutionary consistency constraint training and generating consistency loss values ​​includes: In each training iteration, the prediction step interval corresponding to the windowed reference computing power state evolution sequence is determined, and multi-step continuous state segments of the prediction computing power state evolution sequence are extracted within the prediction step interval. The multi-step continuous state segments are aligned one by one with the reference computing power state evolution sequence within the corresponding time window according to the time step order to form a multi-step state pairing sequence. For each pair of time-step aligned predicted computing power states and reference computing power states, perform state-by-state dimension difference calculation to obtain the state deviation vector for the corresponding time step; perform norm calculation on the state deviation vector in the state dimension to generate the state deviation scalar for the corresponding time step; wherein, performing state-by-state dimension difference calculation means that for each time step aligned predicted computing power state and reference computing power state, take the corresponding values ​​under the same state dimension index and perform subtraction operation to generate a state deviation vector containing the difference of each state dimension, which is used to characterize the degree of deviation between the predicted state and the reference state in each state dimension; The state deviation scalars at each time step are cumulatively weighted according to the time step sequence, with higher weights applied to the state deviation scalars corresponding to more distant prediction time steps, resulting in a multi-step evolution consistency cumulative value. This multi-step evolution consistency cumulative value is determined as the consistency loss value, which is then input into the model parameter update process to constrain the evolution consistency of the predicted computing power state evolution sequence across multiple time scales. Specifically, the consistency loss value is input into the model parameter update process as follows: In each training iteration, the consistency loss value is introduced as a constraint term into the model parameter update calculation. During parameter updates, the update amplitudes of the task condition state matrix, dual-timescale state update gating, and cross-step residual state advancement related parameters are modulated based on the consistency loss value, so that the evolution trajectory of the predicted computing power state evolution sequence across multiple consecutive time steps converges to the reference computing power state evolution sequence, thereby constraining its evolution consistency across multiple time scales.

[0029] In this embodiment, the step of iteratively updating the parameters of the improved S4 state-space sequence model based on the consistency loss value and the scheduling target supervision signal to obtain the convergent model parameter set includes: In each training iteration, the consistency loss value obtained from the multi-step evolution consistency constraint training and the scheduling supervision loss value calculated from the scheduling target supervision signal are obtained respectively. The consistency loss value and the scheduling supervision loss value are weighted and fused according to a preset weight ratio to generate a joint training loss value. Based on the joint training loss value, reverse update calculations are performed on the parameters related to the generation of the task conditional state matrix, the parameters related to the gating of the dual time-scaled state update, and the parameters related to the propagation of the step residual state in the improved S4 state-space sequence model. In this embodiment, in each training iteration, the joint training loss value is used as the optimization target. The gradient information corresponding to the parameters of the improved S4 state space sequence model is calculated. Based on the gradient information, the parameters related to the generation of the task condition state matrix, the parameters related to the gating of the dual time-scale state update, and the parameters related to the advancement of the step residual state are modified respectively, so that all kinds of parameters are updated synchronously in the direction of reducing the joint training loss value. The parameter update step is repeated in several consecutive training iterations, and the corresponding joint training loss value is recorded after each iteration. When the change in the joint training loss value is lower than a preset threshold during a preset number of consecutive training iterations, the model parameters are determined to have reached a convergent state, parameter updates are stopped, and the model parameters obtained at the time of stopping updates are determined as the convergent model parameter set.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a computing power scheduling environment with concurrent execution of multiple tasks. In this environment, a large number of computing tasks are continuously submitted, with significant differences in computational scale, parallelism, and latency constraints, resulting in a continuously changing load state for computing resources. Traditional scheduling methods in such environments are prone to problems such as accumulated prediction offsets of computing power state, large fluctuations in scheduling results, and uneven resource utilization, making it difficult to maintain stable performance during continuous scheduling.

[0031] In the aforementioned environment, data on tasks to be scheduled and data on computing resource operation are continuously collected and input into the method flow of this invention. The data on tasks to be scheduled includes parameters such as task computation scale, task parallelism, and task latency constraints. The data on computing resource operation includes parameters such as computing load, available computing capacity, and resource utilization. Through a unified data preprocessing process, standardized task sequences and standardized computing power state sequences are constructed, providing a consistent data foundation for subsequent modeling. Based on the standardized task sequences, task computation load characteristics and task latency constraint characteristics are extracted to form a task feature sequence. Based on the standardized computing power state sequences, computing load characteristics and computing power availability characteristics are extracted to form a computing power feature sequence. The task feature sequences and computing power feature sequences are simultaneously spliced ​​and associated with each other on a unified timeline to generate a task-computing power joint state sequence.

[0032] The task-computing power joint state sequence is input into an improved state-space sequence model. Internally, the model generates a task-conditional state matrix based on task characteristics, modulating the state transition structure according to changes in task load. A dual-timescale state update gating mechanism distinguishes between rapid state changes caused by sudden load changes and slow state changes caused by computing power capacity evolution during state advancement. A step-residual state advancement mechanism introduces historical state influences and applies time-step interval decay constraints during continuous state updates, constructing a stable state evolution path. Through these mechanisms, the model outputs a predicted computing power state evolution sequence, used to describe the evolution trend of computing power resources across multiple time steps.

[0033] Multi-step evolutionary consistency constraints are introduced during the model training phase. A reference computing power state evolution sequence is constructed based on a standardized computing power state sequence, and the predicted computing power state evolution sequence is aligned with the reference computing power state evolution sequence across multiple consecutive time steps. A state deviation vector is generated by differential calculation along the state dimension, and the state deviation vector is accumulated by norm calculation and time step weighting to form a consistency loss value. The consistency loss value and the scheduling target supervision signal jointly participate in the model parameter update process, performing joint updates on parameters related to task condition state matrix generation, parameters related to dual-timescale state update gating, and parameters related to cross-step residual state advancement. The training process continues until the magnitude of the joint training loss change meets the convergence criterion, forming a converged model parameter set.

[0034] During the inference phase, the convergent model parameter set is fixed. The task-computing power joint state sequence is input into the model to execute forward state advancement, outputting the task-computing power scheduling mapping result. By observing the scheduling effect through continuous scheduling operation, and comparing it with traditional training methods that do not introduce multi-step evolutionary consistency constraints and dual-timescale state update mechanisms, significant differences can be observed. The method of this invention maintains a high degree of consistency between the predicted and actual computing power state during continuous scheduling, significantly reduces the fluctuation amplitude of computing power load, makes the task completion delay distribution more concentrated, and tends to balance resource utilization. No significant state prediction drift phenomenon occurs during scheduling, improving the stability of continuous scheduling.

[0035] To verify the feasibility of this invention in practice, it was applied to a multi-task concurrent computing power scheduling environment. The results of the traditional single-step supervised training method and the method of this invention were compared under the same input conditions. Indicators such as computing power load fluctuation, prediction state deviation, task completion delay distribution, and resource utilization balance were recorded. The comparison results show that after introducing a task condition state matrix, dual-timescale state update gating, and multi-step evolution consistency constraints, the computing power state prediction deviation is significantly reduced, the scheduling results remain stable during continuous operation, and the smoothness of the resource utilization curve is improved. This verifies the practical value of the method of this invention in complex scheduling scenarios.

[0036] Table 1: Comparison of Multi-Task Concurrent Computing Power Scheduling Results

[0037] As can be seen from Table 1, when using the traditional training method, the computing load exhibits obvious fluctuations during continuous scheduling, with a high peak-to-valley difference, reflecting the uneven distribution of computing resources across different time steps. The method of this invention compresses the peak-to-valley difference of computing load under the same scheduling environment, resulting in a smoother computing load change curve. This indicates that the modeling method based on the task condition state matrix and the dual-time-scale state update mechanism can effectively suppress the impact of load mutations on the overall computing power state.

[0038] Regarding the prediction bias of multi-step computing power state, traditional training methods show significant error accumulation during continuous prediction, with the predicted state gradually deviating from the actual operating state. The method of this invention introduces multi-step evolution consistency constraints, which ensures that the evolution trajectory of the predicted computing power state remains consistent with the evolution of the reference computing power state over multiple consecutive time steps. The prediction bias is significantly reduced, indicating that the model's characterization of the evolution law of computing power state is more stable and the long-sequence prediction capability is significantly enhanced.

[0039] From the indicators of task completion delay dispersion and the length of the continuous scheduling stable interval, it can be observed that traditional methods, in multi-task concurrent scenarios, result in a dispersed distribution of task completion times and significant fluctuations in scheduling performance during operation. The method of this invention, under the same conditions, makes the distribution of task completion delays more concentrated and significantly extends the continuous stable scheduling interval, indicating that the model can balance the evolution of computing power status and scheduling objective constraints during the scheduling decision generation process, reducing unnecessary scheduling oscillations.

[0040] In terms of the consistency score between the computing resource utilization balance index and the scheduling mapping result, the method of this invention shows a significant advantage over the traditional method. The improved computing resource utilization balance index reflects a more reasonable allocation of computing resources at different nodes and time steps. The improved consistency score of the scheduling mapping result indicates that the model output results maintain a high degree of consistency under similar task-computing power joint state input, which is beneficial to the scheduling reliability in a continuous operating environment.

[0041] The comprehensive state evolution trajectory fitting error index can further verify the above conclusions. Traditional training methods have a large fitting error for the computing power state evolution path, which is difficult to accurately reflect the real operation trend. The method of this invention significantly reduces the fitting error of the state evolution trajectory through step residual state advancement and consistency constraint training, indicating that the evolution path of the model in the state space is closer to the actual computing power operation law, and verifying the effectiveness and engineering applicability of this invention in complex computing power scheduling scenarios.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for training a task model for computing power scheduling, characterized in that, Includes the following steps: Collect data on tasks to be scheduled and computing resource operation data, and perform preprocessing on the data to be scheduled and computing resource operation data to generate standardized task sequences and standardized computing power status sequences; Based on the standardized task sequence, task computation load characteristics and task latency constraint characteristics are extracted to generate a task feature sequence. Based on the standardized computing power status sequence, computing power load characteristics and computing power availability characteristics are extracted to generate a computing power feature sequence. Perform time-synchronous splicing and association annotation on the task feature sequence and the computing power feature sequence to generate a joint task-computing power state sequence; The task-computing power joint state sequence is input into the improved S4 state space sequence model, which includes three parts: task condition state matrix generation, dual time-scaled state update gating, and step residual state advancement, and outputs the predicted computing power state evolution sequence. A windowed reference computing power state evolution sequence is constructed based on the standardized computing power state sequence. Multi-step evolution consistency constraint training is performed on the predicted computing power state evolution sequence and the reference computing power state evolution sequence to generate a consistency loss value. Based on the consistency loss value and the scheduling target supervision signal, the parameters of the improved S4 state space sequence model are iteratively updated to obtain the convergent model parameter set. Based on the convergent model parameter set, the task-computing power joint state sequence is inferred and calculated, and the task-computing power scheduling mapping result is output.

2. The task model training method for computing power scheduling according to claim 1, characterized in that, The task data to be scheduled includes task submission timestamps, task computation scale parameters, task parallelism parameters, and task latency constraint parameters, and an original task sequence is constructed according to the task submission order; the computing power resource operation data includes computing load parameters, available computing capacity parameters, and resource utilization parameters under the corresponding timestamp of the computing power node, and an original computing power status sequence is constructed according to the time order; the preprocessing includes performing timestamp alignment processing on the original task sequence and the original computing power status sequence to form a time-consistent data sequence; and performing missing value imputation and outlier removal processing on the time-aligned data sequence. The original task sequence after processing is subjected to dimensional normalization to generate a standardized task sequence. The original computing power state sequence is subjected to dimensional normalization to generate a standardized computing power state sequence. 3.The method of claim 2, wherein, The generation process of the task feature sequence and computing power feature sequence includes: For each task time step in the standardized task sequence, the task computational load intensity value is calculated based on the task computational scale parameter and the task parallelism parameter. The task computational load intensity values ​​are arranged in chronological order to form a task computational load characteristic sequence. For each task time step in the standardized task sequence, the remaining schedulable time window length of the task is calculated based on the task delay constraint parameters and the corresponding task submission timestamp. The remaining schedulable time window lengths of the task are arranged in chronological order to form a task delay constraint feature sequence. The task computation load feature sequence and the task latency constraint feature sequence are concatenated along the feature dimension to generate the task feature sequence; For each time step in the standardized computing power state sequence, the computing power load occupancy ratio is calculated based on the computing power node's load parameters and available computing capacity parameters, and arranged in chronological order to form a computing power load characteristic sequence. For each time step in the standardized computing power status sequence, the computing power availability index value is calculated based on the available computing capacity parameter and the resource utilization rate parameter, and arranged in chronological order to form a computing power availability feature sequence. The computing load feature sequence and the computing availability feature sequence are concatenated along the feature dimension to generate the computing power feature sequence.

4. The task model training method for computing power scheduling according to claim 3, characterized in that, The process of generating the task-computing power joint state sequence includes: On a unified time axis, determine the corresponding time step index of the task feature sequence and the computing power feature sequence, and match the task feature vector and computing power feature vector under the same time step index one by one. The matching task feature vector and computing power feature vector are concatenated along the feature dimension to generate a joint feature vector for the corresponding time step. Add a time step identifier and a task identifier to each joint feature vector to form a joint state term that includes time sequence information and task association information; Arrange the joint state items of each time step in the order of time step index to generate the task-computing power joint state sequence.

5. The task model training method for computing power scheduling according to claim 4, characterized in that, The steps for generating the task condition state matrix in the improved S4 state-space sequence model include: At each time step, the task feature vector and computing power feature vector of the corresponding time step are read from the task-computing power joint state sequence, and the task feature vector is used as the conditional input vector. Based on the conditional input vector, perform linear mapping and nonlinear transformation to generate task conditional embedding vectors; Based on the task condition embedding vector, perform element-wise modulation operation on the preset basic state matrix parameters to generate a task condition state matrix corresponding to the current task features; Structural constraints are imposed on the task condition state matrix, maintaining a fixed order in the state dimension and allowing it to change with the task conditions in the time step dimension. The generated task condition state matrix is ​​then used as the state transition matrix for the current time step.

6. The task model training method for computing power scheduling according to claim 5, characterized in that, The dual-time-scaled state update gating steps in the improved S4 state-space sequence model include: At each time step, based on the task-computing power joint state vector of the current time step, a fast time scale update control vector and a slow time scale update control vector are generated respectively. Based on the fast timescale update control vector, high-frequency update calculation is performed on the state increment under the action of the task condition state matrix corresponding to the current time step to obtain the fast timescale state update component, which characterizes the transient computing power state response caused by the change in task load. Based on the slow time-scaled update control vector, a low-frequency smooth update calculation is performed on the accumulated state representation of the previous time step to obtain the slow time-scaled state update component, which characterizes the gradual evolution process of computing power resource capacity and occupancy level. Gating coefficient constraints are applied to the fast time-scaled state update component and the slow time-scaled state update component respectively, so that the two types of update components participate in the state update in the same time step according to a preset ratio; The fast timescale state update component and the slow timescale state update component after gating constraints are weighted and fused to generate the comprehensive state update result for the current time step. The comprehensive state update result is used as the input state for the step residual state advancement.

7. The task model training method for computing power scheduling according to claim 6, characterized in that, The step-by-step residual state advancement step in the improved S4 state-space sequence model includes: At each time step, the candidate state representation for the current time step is generated based on the comprehensive state update result of the dual-time-scaled state update gate output. Read the historical state representation from the task-computing power joint state sequence, which is at least one time step away from the current time step, and use the historical state representation as the step residual term; Apply time step interval weight decay to the step residual term; The cross-step residual term after weight decay is added to the candidate state representation of the current time step element by element to generate the cross-step residual fused state representation; Using the step residual fusion state representation as the initial state input for the next time step, the process is continuously advanced along the time step direction to form a predictive computing power state evolution sequence.

8. The task model training method for computing power scheduling according to claim 7, characterized in that, The steps of performing multi-step evolutionary consistency constraint training and generating consistency loss values ​​include: In each training iteration, the prediction step interval corresponding to the windowed reference computing power state evolution sequence is determined, and multi-step continuous state segments of the prediction computing power state evolution sequence are extracted within the prediction step interval. The multi-step continuous state segments are aligned one by one with the reference computing power state evolution sequence within the corresponding time window according to the time step order to form a multi-step state pairing sequence. For each pair of time-step aligned predicted computing power states and reference computing power states, perform state-by-state dimension difference calculation to obtain the state deviation vector for the corresponding time step; perform norm calculation on the state deviation vector in the state dimension to generate the state deviation scalar for the corresponding time step. The state deviation scalars of each time step are cumulatively weighted according to the time step sequence, with higher weights applied to the state deviation scalars corresponding to more distant prediction time steps, to obtain the cumulative value of multi-step evolution consistency. The cumulative value of multi-step evolution consistency is determined as the consistency loss value, and the consistency loss value is input into the model parameter update process to constrain the evolution consistency of the prediction computing power state evolution sequence on multi-step time scales.

9. A task model training method for computing power scheduling according to claim 8, characterized in that, The steps of iteratively updating the parameters of the improved S4 state-space sequence model based on the consistency loss value and the scheduling target supervision signal to obtain the convergent model parameter set include: In each training iteration, the consistency loss value obtained from the multi-step evolution consistency constraint training and the scheduling supervision loss value calculated from the scheduling target supervision signal are obtained respectively. The consistency loss value and the scheduling supervision loss value are weighted and fused according to a preset weight ratio to generate a joint training loss value. Based on the joint training loss value, reverse update calculations are performed on the parameters related to the generation of the task conditional state matrix, the parameters related to the gating of the dual time-scaled state update, and the parameters related to the propagation of the step residual state in the improved S4 state-space sequence model. The parameter update step is repeated in several consecutive training iterations, and the corresponding joint training loss value is recorded after each iteration. When the change in the joint training loss value is lower than a preset threshold during a preset number of consecutive training iterations, the model parameters are determined to have reached a convergent state, parameter updates are stopped, and the model parameters obtained at the time of stopping updates are determined as the convergent model parameter set.

Images