Simulation resource peak-shaving allocation and scheduling system based on timing optimization

Abstract

This invention discloses a simulation resource staggered allocation and scheduling system based on time-series optimization, belonging to the field of simulation computing cluster resource scheduling technology. Specifically, it includes a time-series-aware resource module, a scheduling task classification module, a staggered planning and management module, a task scheduling decision module, and a resource elastic scheduling module. This invention achieves sensitivity quantification by constructing a multi-dimensional feature set, classifying tasks as time-sensitive or insensitive. Simultaneously, it implements precise staggered scheduling for time-insensitive tasks, fully utilizing off-peak periods and time-of-use pricing advantages to effectively reduce node load peaks and improve resource utilization. Furthermore, it combines time-series load prediction to construct a multi-objective optimization scheduling model, solving for the optimal task-node allocation scheme and scheduling sequence, while simultaneously performing feasibility verification and dynamic monitoring and adjustment.

Description

Technical Field

[0001] This invention relates to the field of simulation computing cluster resource scheduling technology, and in particular to a simulation resource staggered allocation and scheduling system based on time-series optimization. Background Technology

[0002] Simulation technology, as a core supporting means for scientific research innovation, engineering design, and industrial verification, has been widely used in aerospace, automobile manufacturing, biomedicine, electronic information and other fields. With the development of simulation tasks towards scale and complexity, the resource scale of simulation clusters continues to expand, and resource scheduling efficiency has become a key factor restricting the execution efficiency of simulation tasks, system operating costs and task execution quality.

[0003] Currently, simulation resource scheduling schemes are mainly divided into two categories: static allocation strategies and dynamic scheduling strategies. Static allocation strategies use a fixed resource allocation method, which cannot be adjusted according to changes in resource load and task requirements, resulting in low resource utilization and severe task execution delays. While dynamic scheduling strategies can adjust resources based on real-time load, they still have the following prominent technical defects, making it difficult to meet the needs of practical applications: The off-peak allocation mechanism is imperfect: it does not distinguish the time-sensitive attributes of simulation tasks and adopts a uniform scheduling strategy for all tasks. This cannot meet the low-latency execution requirements of time-sensitive tasks (such as real-time simulation and urgent verification tasks) nor make full use of the delay characteristics of time-insensitive tasks (such as batch simulation and offline computing tasks). It cannot achieve off-peak resource load, resulting in resource load imbalance during peak and valley periods and high operating costs.

[0004] Traditional scheduling decisions often focus on a single objective (such as the shortest task completion time or the highest resource utilization rate) without comprehensively considering multiple dimensions such as resource utilization, task execution delay, operating costs, and energy consumption. This results in poor overall efficiency of the scheduling solution and an inability to adapt to the differentiated needs of different business scenarios.

[0005] Weak elasticity adjustment capability: The resource scheduling lacks a closed-loop feedback mechanism. When the node load is abnormal, the task execution fails (such as timeout or failure), or the task requirements change, it cannot respond quickly and adjust the scheduling scheme, which can easily lead to task execution interruption, resource waste, and affect the operational stability of the entire simulation cluster.

[0006] Therefore, there is an urgent need for a simulation resource scheduling scheme that can accurately allocate resources during off-peak periods, perform multi-objective intelligent optimization, and dynamically adjust resources to achieve efficient utilization of simulation resources, cost savings, and stable task execution. Summary of the Invention

[0007] This invention aims to solve the technical problems of traditional simulation resource scheduling schemes, such as lack of time-series awareness, insufficient off-peak allocation, lack of multi-objective optimization, and weak elastic adjustment capability. It provides a simulation resource off-peak allocation and scheduling system based on time-series optimization. Through time-series data awareness, task classification off-peak, intelligent multi-objective decision-making, and elastic closed-loop scheduling, it achieves efficient, low-cost, and highly stable scheduling of simulation resources, thereby improving the execution efficiency of simulation tasks and the overall operational benefits of the system.

[0008] The objective of this invention can be achieved through the following technical solution: a simulation resource staggered allocation and scheduling system based on time-series optimization, including a resource dynamic management center, a time-series-aware resource module, a scheduling task classification module, a staggered planning and management module, a task scheduling decision module, and a resource elastic scheduling module; The time-aware resource module is used to collect real-time load time-series data of each computing node, historical load time-series dataset of the node, and task information of the simulation tasks to be scheduled, and to construct a multi-dimensional feature set containing the time-series characteristics of node resource status and the time-series constraints of tasks. The scheduling task classification module is used to retrieve a multi-dimensional feature set, perform discrimination processing on the sensitivity quantification value S obtained from the analysis, and output time-sensitive tasks and time-insensitive tasks. The peak-shifting planning management module is used to perform peak-shifting planning analysis on tasks marked as time-insensitive, obtain candidate peak-shifting execution windows, comprehensively evaluate and optimize simulation tasks with multiple candidate peak-shifting execution windows, and output the final peak-shifting execution window and task scheduling pre-planning results. The task scheduling decision module is used to predict the load forecast value of each node within a preset time step based on the time-series load prediction model. Combined with the pre-planning results of task scheduling, it constructs a multi-objective optimization scheduling model, outputs task-node allocation schemes and scheduling sequences, and performs subsequent feasibility verification and execution monitoring scheduling operations.

[0009] Preferably, the construction and analysis process of the multidimensional feature set is as follows: S1: Deploy monitoring agents for each computing node in the simulation cluster and establish data acquisition channels; S2: Collect resource load metadata of each node in the simulation cluster through the data acquisition channel at a preset sampling period, and add timestamps to form real-time load time-series data including CPU, memory and GPU utilization. S3: After cleaning and formatting the collected real-time load time-series data, store the processed real-time load time-series data in a distributed time-series database. For each computing node, the distributed time-series database continuously stores its long-term historical load change curve with time as the axis, forming a node historical load time-series dataset. S4: Parse the simulation task submitted by the user and extract the task information of the simulation task, including task attributes, resource requirements, execution time limit and timing preference; S5: Based on the collected node load time-series data and task information, extract time-series features and construct node resource status time-series features and task time-series constraint features; S6: Integrate the temporal features of node resource status, task attributes, resource requirements, and task temporal constraints to form a multidimensional feature set in a unified format.

[0010] Preferably, the classification and analysis process for time-sensitive tasks and time-insensitive tasks is as follows: For each simulation task to be scheduled in the multidimensional feature set, extract the sensitivity quantification value S from the temporal preference field of the simulation task, and compare the sensitivity quantification value S with the preset sensitivity classification threshold: If the sensitivity quantification value S is greater than or equal to the preset sensitivity classification threshold, it is determined to be a time-sensitive task; if the sensitivity quantification value S is less than the preset sensitivity classification threshold, it is determined to be a time-insensitive task.

[0011] Preferably, the analysis process for the sensitivity quantification value S is as follows: Based on the multidimensional feature set, task time-series constraint features for task classification and node load time-series features and time-of-use electricity price data for window planning are obtained; For all tasks that have been marked as time-insensitive, extract the key constraint parameters for peak-shifting planning for each time-insensitive task. Based on time-of-use electricity price data, a task feasibility domain, an electricity price optimization domain, and a load optimization domain are constructed. The intersection of the feasible region of the task, the preferred region of electricity price, and the preferred region of load is calculated to obtain one or more continuous time intervals, which are marked as candidate peak-shifting execution windows. The simulation task that generates multiple candidate staggered execution windows is comprehensively evaluated and optimized, and the final staggered execution window is output. Time-sensitive tasks, time-insensitive tasks, and their corresponding off-peak execution windows are structured and encapsulated to form pre-planning results for task scheduling.

[0012] Preferably, the task feasible domain represents the time range [Tinitial, Tend] from the task submission time Tinitial to the latest completion deadline Tend; the electricity price preferred domain represents the set of low electricity price periods in the electricity price time series curve {Tprice|P(t)<electricity price threshold}; and the load preferred domain represents the set of low load periods for nodes {Tlow|t<Llow}.

[0013] Preferably, the comprehensive evaluation and optimization analysis process is as follows: Obtain cost savings estimates and load smoothing contributions for simulation tasks with multiple candidate off-peak execution windows; The preset proportional coefficients a1 and a2 of the cost saving estimate and the load smoothing contribution are retrieved. Since a1 and a2 are both greater than zero, the comprehensive score is calculated based on the cost saving estimate × a1 + load smoothing contribution × a2. The candidate staggered execution window corresponding to the maximum value of the comprehensive score is set as the final staggered execution window.

[0014] Preferably, the analysis process of the task-node allocation scheme and scheduling sequence is as follows: SS1: Obtain the real-time load data of each node over the most recent N (N>0) time steps to form the input sequence; SS2: Input the input sequence into the pre-set time series load prediction model and output the load prediction values ​​of each node for the next H (H>0) time steps, forming a set of node load prediction curves Kji, where j is the node index and i is the future time point; SS3: Combining the pre-planning results of task scheduling with the time-series load prediction results, construct a multi-objective optimization scheduling model and clarify the optimization objectives and constraints; SS4: Use existing improved genetic algorithms to solve the multi-objective optimization scheduling model and output the optimal scheduling solution; SS5: Analyzes the optimal scheduling solution and generates a structured task-node allocation scheme and scheduling sequence.

[0015] Preferably, it also includes SS6: to perform feasibility verification on the generated task-node allocation scheme and scheduling sequence. The verification includes: whether the node resources meet the task requirements, whether the task execution meets the timing constraints, whether the node load exceeds the safety threshold, and whether the running cost meets the expectations. If the verification fails, return to SS4 to readjust the genetic algorithm parameters and solve until a feasible scheme is generated. SS7: Synchronize the verified task-node allocation scheme and scheduling sequence to the resource elastic scheduling module; SS8: Continuously monitors the execution information of the scheduling scheme, including node load prediction deviation and task execution delay results (including abnormal and normal). The execution information is judged. If the node load prediction deviation exceeds the preset threshold or the task execution delay result is abnormal, SS4 is triggered to resolve and update the task-node allocation scheme and scheduling sequence.

[0016] The beneficial effects of this invention are as follows: This invention achieves time-series load prediction and accurate task time-series classification by collecting and analyzing the time-series characteristics of resource load and tasks. It avoids the blindness of traditional scheduling, improves the foresight and rationality of resource allocation, and enables precise peak-shifting scheduling for time-insensitive tasks. It makes full use of the advantages of low load periods and time-of-use electricity pricing, effectively reduces node load peaks, improves resource utilization, and enables differentiated and refined scheduling of different types of tasks to meet diverse task execution needs.

[0017] This invention also accurately predicts future load changes of nodes based on a time-series load prediction model, combines the pre-planning results of task scheduling to construct a multi-objective optimization scheduling model, and solves the optimal solution through an improved genetic algorithm, taking into account both task execution quality and system operation efficiency, adapting to the needs of different business scenarios, and having a real-time monitoring and dynamic adjustment mechanism, which can quickly respond to scenarios such as abnormal node load and task execution failure, ensuring the continuity and stability of task execution. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings; Figure 1 This is a flowchart of the system of the present invention; Figure 2 This is a reference diagram for analyzing the task classification module of this invention; Figure 3 This is a reference diagram for analyzing the task scheduling decision module of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments; Example 1: Please refer to Figures 1 to 3As shown, this invention is a simulation resource staggered allocation and scheduling system based on time-series optimization, including a resource dynamic management center, a time-series-aware resource module, a scheduling task classification module, a staggered planning management module, a task scheduling decision module, and a resource elastic scheduling module. The resource dynamic management center and the time-series-aware resource module have a bidirectional communication connection, and the resource dynamic management center has a unidirectional communication connection with both the scheduling task classification module and the task scheduling decision module. The scheduling task classification module has a unidirectional communication connection with the staggered planning management module, the staggered planning management module has a unidirectional communication connection with the resource dynamic management center, and the task scheduling decision module has a unidirectional communication connection with the resource elastic scheduling module. The time-aware resource module is used to collect real-time load time-series data of each computing node, historical load time-series datasets of nodes, and task information of simulation tasks to be scheduled. It constructs a multi-dimensional feature set containing time-series characteristics of node resource status and time-series constraints of tasks, specifically including: S1: Deploy monitoring agents for each computing node in the simulation cluster and establish data acquisition channels; S2: Collect resource load metadata of each node in the simulation cluster through the data acquisition channel at a preset sampling period, and add timestamps to form real-time load time-series data containing information such as CPU, memory, and GPU utilization. S3: After cleaning (such as removing outliers and filling missing points) and formatting the collected real-time load time-series data, store the processed real-time load time-series data in a distributed time-series database. For each computing node, the distributed time series database continuously stores its long-term historical load change curve with time as the axis, forming a node historical load time series dataset. This dataset is not only used for real-time monitoring and display, but more importantly, it serves as a training sample for the time series prediction model in the subsequent task scheduling decision module. S4: Parse the simulation task submitted by the user and extract the task information of the simulation task, including task attributes, resource requirements, execution time limit and timing preference; Task attributes include task type, computing scale, etc.; resource requirements include number of cores, memory, etc.; execution time limits include latest start time, latest finish time, etc.; timing preferences include allowed delay duration, priority execution period, etc. S5: Based on the collected node load time-series data and task information, perform time-series feature extraction: Extract time-series features such as peak periods, off-peak periods, and fluctuation cycles from the historical load of nodes, and extract time-series features such as execution time limit sensitivity and latency tolerance from task information to construct the time-series features of node resource status and task time-series constraints. S6: Integrate data such as node resource status time-series characteristics, task attributes, resource requirements, and task time-series constraint characteristics to form a multi-dimensional feature set in a unified format, and send the multi-dimensional feature set to the resource dynamic management center. The task classification module retrieves a multi-dimensional feature set, performs discrimination processing on the analyzed sensitivity quantification value S, and outputs time-sensitive tasks and time-insensitive tasks, specifically including: For each simulation task to be scheduled in the multidimensional feature set, extract the sensitivity quantification value S from the temporal preference field of the simulation task, and compare the sensitivity quantification value S with the preset sensitivity classification threshold: If the sensitivity quantification value S is greater than or equal to the preset sensitivity classification threshold, it is determined to be a time-sensitive task. Such tasks have low tolerance for scheduling delays and need to be allocated resources immediately and executed as soon as possible, such as online interactive simulation or emergency computing tasks. That is, time-sensitive tasks are directly marked as immediate scheduling type and no off-peak planning is performed. If the sensitivity quantification value S is less than the preset sensitivity classification threshold, it is determined to be a time-series non-sensitive task; The process involves extracting completed simulation tasks from historical scheduling logs, using the ratio of the actual waiting time of the simulation task to the total task time as the actual latency rate, and combining the task completion status (such as whether it has timed out) to label each historical task with a true sensitivity label (0 indicates extremely insensitive, 1 indicates extremely sensitive), constructing a training sample set, and building a sensitivity mapping model based on the training sample set. The collected real-time load time series data and task information are normalized to unify indicators with different dimensions and value ranges into the [0, 1] interval; The normalization process includes linear normalization, logarithmic normalization, and one-hot encoding. Linear normalization: For indicators with clear upper and lower limits (such as priority levels), use (x-min) / (max-min); Log normalization: For indicators with large distribution spans (such as estimated execution time), use log(x) / log(max); One-hot encoding: For categorical indicators (such as task type labels), convert them into multiple 0 / 1 binary features; The normalized feature vector is input into the sensitivity mapping model, and the predicted value output by the sensitivity mapping model is directly obtained as the sensitivity quantification value S.

[0021] Example 2: The off-peak planning management module is used to perform off-peak planning analysis on tasks marked as time-insensitive, obtain candidate off-peak execution windows, comprehensively evaluate and optimize simulation tasks with multiple candidate off-peak execution windows, and output the final off-peak execution window and task scheduling pre-planning results, specifically including: Based on the multidimensional feature set, task time-series constraint features for task classification and node load time-series features and time-of-use electricity price data for window planning are obtained; For all tasks marked as time-insensitive, extract the key constraint parameters for peak-shifting planning for each time-insensitive task: Task submission timestamp: The initial time T when the task entered the scheduler; Latest completion deadline: The deadline T that a user-specified or default task must complete; Tolerable delay duration: calculated by the difference between T_stop and T_initial; Based on time-of-use electricity price data, the electricity cost price for each time period within a future cycle (e.g., the next 24 hours) is obtained, forming an electricity price time series curve P(t), where t is a time point, and the curve marks low electricity price periods (e.g., 23:00-7:00 the next day) and high electricity price periods; Historical load data of each node is read from a distributed time-series database. Periodic analysis algorithms (such as STL time series decomposition) are used to identify the daily periodic resource load trough L (e.g., 02:00-06:00 in the morning) and the trend inflection point of load transition from peak to trough. Construct the task feasibility domain, the electricity price optimization domain, and the load optimization domain; The feasible region of a task represents the time range [initial T, ending T] from the task submission time T to the latest completion deadline T. The preferred electricity price domain represents the set of low electricity price periods in the electricity price time series curve {T price | P(t) < electricity price threshold}; The load preference domain represents the set of periods when a node's load is at its lowest {T_low|t<L_low}; The intersection of the feasible region of the task, the preferred region of electricity price, and the preferred region of load is calculated to obtain one or more continuous time intervals, which are marked as candidate peak-shifting execution windows. The simulation task with multiple candidate staggered execution windows was comprehensively evaluated and optimized. The analysis process is as follows: Obtain cost savings estimates and load smoothing contributions for simulation tasks with multiple candidate off-peak execution windows; Cost savings projection: Calculate the electricity cost savings that can be achieved by executing the task in different candidate windows compared to executing it during peak hours; Load smoothing contribution: Retrieve a preset baseline load curve and calculate the variance Vz of the baseline load curve within a single candidate off-peak execution window and a period before and after it (e.g., twice the window length). The larger the variance, the more severe the load fluctuation. The resource requirements of the simulation task are superimposed on the baseline load curve to obtain the corrected load curve, and the variance Vn of the same candidate off-peak execution window is calculated. The load smoothing contribution is calculated based on (variance Vz - variance Vn) / variance Vz; The preset proportional coefficients a1 and a2 of the cost saving estimate and the load smoothing contribution are retrieved. Since a1 and a2 are both greater than zero, the comprehensive score is calculated based on the cost saving estimate × a1 + load smoothing contribution × a2. The candidate staggered execution window corresponding to the maximum value of the comprehensive score is set as the final staggered execution window. Time-sensitive tasks, time-insensitive tasks, and their corresponding off-peak execution windows are structured and encapsulated to form task scheduling pre-planning results; The task scheduling decision module is used to predict the load forecast values ​​of each node within a preset time step based on the time-series load prediction model. Combined with the pre-planning results of task scheduling, it constructs a multi-objective optimization scheduling model, outputs a task-node allocation scheme and scheduling sequence, and performs subsequent feasibility verification and execution monitoring scheduling operations, specifically including: SS1: Obtain the real-time load data of each node over the most recent N (N>0) time steps to form the input sequence; SS2: Input the input sequence into the pre-set time series load prediction model and output the node load prediction values ​​for the next H (H>0) time steps (e.g., one prediction point every 5 minutes in the next 3 hours), forming a set of node load prediction curves Kji, where j is the node index and i is the future time point; SS3: Combining the pre-planning results of task scheduling with the time-series load prediction results, construct a multi-objective optimization scheduling model, and clarify the optimization objectives and constraints: Optimization objectives: Maximize the utilization of simulation cluster resources, minimize task execution latency, and minimize operating costs; Constraints: matching task resource requirements with node available resources; real-time scheduling constraints for time-sensitive tasks; scheduling constraints for time-insensitive tasks within off-peak execution windows; task execution not exceeding the latest completion time; node load not exceeding the safety threshold. SS4: Use existing improved genetic algorithms to solve the multi-objective optimization scheduling model and output the optimal scheduling solution; The execution process of the existing optimization algorithm is as follows: Encoding: The task-node allocation scheme and task execution time sequence are encoded into binary to generate the initial population; Fitness function: Combining the three optimization objectives, a weighted fitness function is designed to quantify the merits of each individual (scheduling scheme); Genetic operations: Adaptive crossover and mutation mechanisms and time constraint repair mechanisms are introduced. The crossover and mutation probabilities are dynamically adjusted according to the population fitness to avoid premature convergence of the algorithm, while ensuring that the generated scheduling scheme meets the time constraints and resource constraints. Elite retention: The best individuals in each generation of the population are retained and directly enter the next generation iteration, improving the solution efficiency and the quality of the best solution; Iteration Termination: When the number of iterations reaches a preset threshold, or the fitness of the optimal solution tends to stabilize, the iteration terminates and the optimal scheduling solution is output. SS5: Analyzes the optimal scheduling solution, generating structured task-node allocation schemes and scheduling sequences, specifically including: For time-sensitive tasks: clearly define the corresponding allocated computing nodes, immediate start time, and resource allocation (such as the number of CPU cores, memory, GPU, etc.); For time-insensitive tasks: clearly define the corresponding allocated computing nodes, the optimal start time within the off-peak execution window, and the resource allocation amount, etc. The execution order and resource release time of all tasks are determined to ensure efficient reuse of node resources and avoid resource conflicts. SS6: Perform feasibility verification on the generated task-node allocation scheme and scheduling sequence. The verification includes: whether the node resources meet the task requirements, whether the task execution meets the time constraints (immediacy of sensitive tasks, off-peak window for non-sensitive tasks), whether the node load exceeds the safety threshold, and whether the running cost meets the expectations. If the verification fails, return to SS4 to readjust the genetic algorithm parameters and solve the problem until a feasible solution is generated. SS7: Synchronize the verified task-node allocation scheme and scheduling sequence to the resource elastic scheduling module for subsequent resource allocation and task scheduling execution. At the same time, back up the scheduling scheme to local storage for comparison and reference during subsequent dynamic adjustments. SS8: Continuously monitor the execution information of the scheduling scheme, including node load prediction deviation and task execution delay results (including abnormal and normal). Judge the execution information. If the node load prediction deviation exceeds the preset threshold or the task execution delay result is abnormal, SS4 is triggered to resolve and update the task-node allocation scheme and scheduling sequence to ensure the real-time performance and effectiveness of scheduling decisions. In summary, by collecting and analyzing the temporal characteristics of resource load and tasks, this system achieves time-series load prediction and accurate task classification, avoiding the blindness of traditional scheduling and improving the foresight and rationality of resource allocation. Simultaneously, it enables precise off-peak scheduling for time-insensitive tasks, fully utilizing off-peak periods and time-of-use pricing advantages to effectively reduce node load peaks and improve resource utilization. This allows for differentiated and refined scheduling of different task types, meeting diverse task execution needs. Based on the time-series load prediction model, it accurately predicts future node load changes. Combined with the pre-planning results of task scheduling, a multi-objective optimization scheduling model is constructed, and an improved genetic algorithm is used to solve for the optimal solution. This model balances task execution quality and system operational efficiency, adapting to different business scenarios and possessing real-time monitoring and dynamic adjustment mechanisms. It can quickly respond to scenarios such as abnormal node load and task execution failures, ensuring the continuity and stability of task execution.

[0022] The threshold is set for comparative analysis of results to determine whether they are good or bad. The value of the threshold is determined by a combination of large-scale model analysis of sample data and human experience. It can also be adjusted appropriately based on seasonal or common-sense influencing factors.

[0023] 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 simulation resource staggered allocation and scheduling system based on time-series optimization, characterized in that, It includes a resource dynamic management center, a time-aware resource module, a scheduling task classification module, a peak-shifting planning management module, a task scheduling decision module, and a resource elastic scheduling module; The time-aware resource module is used to collect real-time load time-series data of each computing node, historical load time-series dataset of the node, and task information of the simulation tasks to be scheduled, and to construct a multi-dimensional feature set containing the time-series characteristics of node resource status and the time-series constraints of tasks. The scheduling task classification module is used to retrieve a multi-dimensional feature set, perform discrimination processing on the sensitivity quantification value S obtained from the analysis, and output time-sensitive tasks and time-insensitive tasks. The peak-shifting planning management module is used to perform peak-shifting planning analysis on tasks marked as time-insensitive, obtain candidate peak-shifting execution windows, comprehensively evaluate and optimize simulation tasks with multiple candidate peak-shifting execution windows, and output the final peak-shifting execution window and task scheduling pre-planning results. The task scheduling decision module is used to predict the load forecast value of each node within a preset time step based on the time-series load prediction model. Combined with the pre-planning results of task scheduling, it constructs a multi-objective optimization scheduling model, outputs task-node allocation schemes and scheduling sequences, and performs subsequent feasibility verification and execution monitoring scheduling operations.

2. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 1, characterized in that, The construction and analysis process of the multidimensional feature set is as follows: S1: Deploy monitoring agents for each computing node in the simulation cluster and establish data acquisition channels; S2: Collect resource load metadata of each node in the simulation cluster through the data acquisition channel at a preset sampling period, and add timestamps to form real-time load time-series data including CPU, memory and GPU utilization. S3: After cleaning and formatting the collected real-time load time-series data, store the processed real-time load time-series data in a distributed time-series database. For each computing node, the distributed time-series database continuously stores its long-term historical load change curve with time as the axis, forming a node historical load time-series dataset. S4: Parse the simulation task submitted by the user and extract the task information of the simulation task, including task attributes, resource requirements, execution time limit and timing preference; S5: Based on the collected node load time-series data and task information, extract time-series features and construct node resource status time-series features and task time-series constraint features; S6: Integrate the temporal features of node resource status, task attributes, resource requirements, and task temporal constraints to form a multidimensional feature set in a unified format.

3. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 1, characterized in that, The classification and analysis process for time-sensitive and time-insensitive tasks is as follows: For each simulation task to be scheduled in the multidimensional feature set, extract the sensitivity quantification value S from the temporal preference field of the simulation task, and compare the sensitivity quantification value S with the preset sensitivity classification threshold: If the sensitivity quantification value S is greater than or equal to the preset sensitivity classification threshold, it is determined to be a time-sensitive task; if the sensitivity quantification value S is less than the preset sensitivity classification threshold, it is determined to be a time-insensitive task.

4. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 2, characterized in that, The analysis process for the sensitivity quantification value S is as follows: Based on the multidimensional feature set, task time-series constraint features for task classification and node load time-series features and time-of-use electricity price data for window planning are obtained; For all tasks that have been marked as time-insensitive, extract the key constraint parameters for peak-shifting planning for each time-insensitive task. Based on time-of-use electricity price data, a task feasibility domain, an electricity price optimization domain, and a load optimization domain are constructed. The intersection of the feasible region of the task, the preferred region of electricity price, and the preferred region of load is calculated to obtain one or more continuous time intervals, which are marked as candidate peak-shifting execution windows. The simulation task that generates multiple candidate staggered execution windows is comprehensively evaluated and optimized, and the final staggered execution window is output. Time-sensitive tasks, time-insensitive tasks, and their corresponding off-peak execution windows are structured and encapsulated to form pre-planning results for task scheduling.

5. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 4, characterized in that, The feasible domain of a task represents the time range [Tinitial, Tend] from the task submission time Tinitial to the latest completion deadline Tend; the preferred electricity price domain represents the set of low electricity price periods in the electricity price time series curve {Tprice|P(t)<electricity price threshold}; the preferred load domain represents the set of low load periods for nodes {Tlow|t<Llow}.

6. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 5, characterized in that, The comprehensive evaluation and optimization analysis process is as follows: Obtain cost savings estimates and load smoothing contributions for simulation tasks with multiple candidate off-peak execution windows; The preset proportional coefficients a1 and a2 of the cost saving estimate and the load smoothing contribution are retrieved. Since a1 and a2 are both greater than zero, the comprehensive score is calculated based on the cost saving estimate × a1 + load smoothing contribution × a2. The candidate staggered execution window corresponding to the maximum value of the comprehensive score is set as the final staggered execution window.

7. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 1, characterized in that, The analysis process of the task-node allocation scheme and scheduling sequence is as follows: SS1: Obtain the real-time load data of each node over the most recent N (N>0) time steps to form the input sequence; SS2: Input the input sequence into the pre-set time series load prediction model and output the load prediction values ​​of each node for the next H (H>0) time steps, forming a set of node load prediction curves Kji, where j is the node index and i is the future time point; SS3: Combining the pre-planning results of task scheduling with the time-series load prediction results, construct a multi-objective optimization scheduling model and clarify the optimization objectives and constraints; SS4: Use existing improved genetic algorithms to solve the multi-objective optimization scheduling model and output the optimal scheduling solution; SS5: Analyzes the optimal scheduling solution and generates a structured task-node allocation scheme and scheduling sequence.

8. The simulation resource staggered allocation and scheduling system based on time-series optimization according to claim 7, characterized in that, It also includes SS6: to perform feasibility verification on the generated task-node allocation scheme and scheduling sequence. The verification includes: whether the node resources meet the task requirements, whether the task execution meets the timing constraints, whether the node load exceeds the safety threshold, and whether the running cost meets the expectations. If the verification fails, it returns to SS4 to readjust the genetic algorithm parameters and solve the problem until a feasible solution is generated. SS7: Synchronize the verified task-node allocation scheme and scheduling sequence to the resource elastic scheduling module; SS8: Continuously monitors the execution information of the scheduling scheme, including node load prediction deviation and task execution delay results (including abnormal and normal). The execution information is judged. If the node load prediction deviation exceeds the preset threshold or the task execution delay result is abnormal, SS4 is triggered to resolve and update the task-node allocation scheme and scheduling sequence.

Images