Infectious disease prevention and control big data ai resource scheduling system and method

Abstract

The present application belongs to the field of medical and health informatization technology, and discloses an infectious disease prevention and control big data AI resource scheduling system and method; by constructing a load temperature field model of a computing power node cluster, the node load intensity is mapped to a temperature distribution, and the temperature gradient between adjacent nodes is calculated to identify hot and cold node pairs. Based on the data dependency topology, the energy consumption cost and time delay penalty of task migration are analyzed, and a multi-path heat diffusion network is constructed combined with the temperature gradient to realize the optimized shunting of tasks from hot spots to multiple cold nodes. The system analyzes the periodic fluctuations and sudden peak characteristics of resource occupation through wavelet transform, dynamically adjusts the task decomposition granularity and load distribution weight, and adapts to changes in resource usage patterns. The present application realizes global heat balance and efficient utilization of computing power resources in epidemic analysis, improves computing efficiency, and enhances the system's ability to respond to sudden computing demands.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, and more specifically, to a big data AI resource scheduling system and method for infectious disease prevention and control. Background Technology

[0002] In the process of big data AI analysis for infectious disease prevention and control, existing resource scheduling technologies face the problem of extremely uneven distribution of computing resources. When an epidemic breaks out suddenly or regionally, the computing tasks carried by key nodes increase exponentially, forming significant "hot spots." CPU utilization often remains above 90%, memory usage approaches a bottleneck, and I / O channels are saturated. At the same time, a large number of edge computing nodes are in a "cold spot" state, with resource utilization below 30%, resulting in an imbalance in the overall utilization of computing resources. Traditional scheduling systems use simple threshold triggering or polling allocation mechanisms, lacking the ability to globally perceive load distribution and accurately identify abnormal areas of "temperature gradients." This leads to delays in the training and inference tasks of key predictive models during peak periods of epidemic data analysis. Existing technologies also generally ignore the data dependency characteristics of tasks. Blind migration results in a large amount of intermediate state data being transmitted in the network, which not only increases energy consumption but also causes significant latency penalties, which is particularly fatal in epidemic monitoring scenarios with high timeliness requirements. Existing methods lack the ability to dynamically perceive and predict resource usage patterns and cannot identify periodic fluctuations and sudden peak characteristics in epidemic data processing, resulting in resource allocation always lagging behind changes in demand. In cross-regional joint prevention and control efforts, computing nodes in different regions have formed "information silos," lacking a collaborative scheduling mechanism. This prevents global resources from flowing efficiently to the front lines of epidemic prevention where computing power is urgently needed. This disparity in computing resources has constrained the training speed and prediction accuracy of key AI models in infectious disease control, and delayed the real-time decision-making process for epidemic situation analysis and intervention measures.

[0003] In view of this, the present invention proposes a big data AI resource scheduling system and method for infectious disease prevention and control to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a big data AI resource scheduling method for infectious disease prevention and control, comprising: Obtain real-time load data and infectious disease epidemic situation data for each computing node; Based on real-time load data, a load temperature field model of the computing node cluster is constructed, which maps the load intensity of the computing nodes to the temperature distribution. By calculating the temperature gradient between adjacent nodes in the load temperature field model, hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold are identified. Obtain the data dependency topology graph of the task to be scheduled, and calculate the transmission energy consumption cost and state synchronization latency penalty of the task migrating between hot nodes and cold nodes based on the data dependency topology graph; Based on transmission energy consumption cost and state synchronization delay penalty, and combined with temperature gradient, a multi-path heat diffusion network for task migration is constructed. The multi-path heat diffusion network describes the diversion path of tasks from hot nodes to multiple cold nodes. Based on the multi-path heat diffusion network and epidemic situation data, a cross-node task decomposition scheme and a spatiotemporal coupled resource reservation strategy are generated. The spatiotemporal coupled resource reservation strategy locks the computing resources of cold nodes in both time and space dimensions. By collecting the resource usage pulse sequence of each node in real time during task execution, wavelet transform is performed on the resource usage pulse sequence to identify the periodic fluctuation pattern and sudden peak characteristics of resource usage. Based on the frequency drift of the periodic fluctuation pattern and the amplitude changes of the sudden peak characteristics, the task decomposition granularity in the task decomposition scheme and the node load allocation weight in the multi-path heat diffusion network are dynamically adjusted. Based on the adjusted task decomposition granularity and node load allocation weights, the resource scheduling system is controlled to perform task migration and resource allocation, thereby achieving thermal balance and efficient utilization of computing resources.

[0005] An infectious disease prevention and control big data AI resource scheduling system, which implements methods for scheduling infectious disease prevention and control big data AI resources, including: The data acquisition module is used to acquire real-time load data and infectious disease epidemic situation data of each computing node; The temperature field modeling module is used to construct a load temperature field model of the computing node cluster based on real-time load data. The hot spot identification module is used to identify hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold by calculating the temperature gradient between adjacent nodes in the load temperature field model. The task analysis module is used to obtain the data dependency topology graph of the task to be scheduled, and calculate the transmission energy consumption cost and state synchronization latency penalty of the task migrating between hot nodes and cold nodes based on the data dependency topology graph. The heat dissipation network construction module is used to construct a multi-path heat dissipation network for task migration based on transmission energy consumption cost and state synchronization delay penalty, combined with temperature gradient. The task decomposition module is used to generate cross-node task decomposition schemes and spatiotemporal coupled resource reservation strategies based on multi-path heat spread network and epidemic situation data. The spatiotemporal coupled resource reservation strategy locks the computing resources of cold nodes in both time and space dimensions. The resource monitoring module is used to collect the resource usage pulse sequence of each node in real time during task execution, perform wavelet transform on the resource usage pulse sequence, and identify the periodic fluctuation pattern and sudden peak characteristics of resource usage. The dynamic adjustment module is used to dynamically adjust the task decomposition granularity in the task decomposition scheme and the node load allocation weight in the multi-path heat diffusion network based on the frequency drift of the periodic fluctuation pattern and the amplitude changes of the sudden peak characteristics. The execution control module is used to control the resource scheduling system to perform task migration and resource allocation based on the adjusted task decomposition granularity and node load allocation weight. The modules are connected via wired and / or wireless means to enable data transmission between them.

[0006] The technical effects and advantages of the big data AI resource scheduling system and method for infectious disease prevention and control of this invention are as follows: This invention achieves a natural and balanced flow of global computing resources, effectively relieving pressure on high-load nodes while fully utilizing the idle resources of low-load nodes, significantly improving overall computing efficiency. During sudden outbreaks of an epidemic, this invention can keenly sense load change trends and deploy computing resources in advance, transforming passive response into proactive prevention, providing more timely and stable computing support for epidemic situation analysis. Through precise task migration strategies, this invention reduces energy consumption and response latency, meeting the stringent real-time requirements of epidemic monitoring and early warning. In complex and ever-changing epidemic environments, this invention can automatically adjust and optimize strategies based on dynamic changes in resource usage patterns without manual intervention, effectively coping with periodic load fluctuations and sudden computing demands. The parallel computing efficiency of this invention is qualitatively improved, accelerating the training of epidemic transmission models and the processing speed of regional risk assessments, providing strong data support for epidemic prevention decision-making. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the big data AI resource scheduling method for infectious disease prevention and control of the present invention; Figure 2 This is a schematic diagram of the big data AI resource scheduling system for infectious disease prevention and control according to the present invention. Detailed Implementation

[0008] 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.

[0009] This application provides an example of a big data AI resource scheduling system and method for infectious disease prevention and control. The executing entities of the system include, but are not limited to, those running the system such as: an epidemic prediction and analysis platform, a medical resource allocation center, an intelligent computing power management system, and a big data processing cluster, which can be considered general computing nodes of this application. The resource scheduling system includes, but is not limited to, at least one of: a cloud load balancing engine, a distributed resource scheduler, and an intelligent task decomposer.

[0010] Please see Figure 1 In this embodiment of the invention, the specific implementation process of the big data AI resource scheduling method for infectious disease prevention and control includes: Real-time load data and infectious disease epidemic situation data are acquired for each computing node. Real-time load data includes key performance indicators such as node CPU utilization, memory utilization, disk I / O throughput, and network I / O throughput, acquired in real-time through a multi-channel data acquisition interface. Infectious disease epidemic situation data records epidemic development trends, regional distribution, transmission rates, and the effectiveness of intervention measures, providing a basis for predicting task demand for computing resource scheduling. Node load data reflects the current usage of computing resources, while epidemic situation data predicts future trends in computing demand. This data provides a comprehensive raw data foundation for subsequent hotspot identification and resource optimization, ensuring the targeted and forward-looking nature of scheduling decisions.

[0011] Based on real-time load data, a load-temperature field model of the computing node cluster is constructed. This model maps node load intensity to temperature distribution, including normalization, weight allocation, and nonlinear mapping of multi-dimensional resource indicators such as CPU, memory, disk, and network. This thermodynamic modeling method intuitively represents high-load nodes as "high-temperature zones" and low-load nodes as "low-temperature zones," facilitating subsequent resource scheduling optimization based on heat diffusion theory. The temperature field model captures both the load status of individual nodes and reflects the load distribution topology of the entire cluster, providing a theoretical basis for hotspot identification and load balancing.

[0012] By calculating the temperature gradient between adjacent nodes in the load temperature field model, hot and cold node pairs with temperature gradients exceeding a preset threshold are identified. First, node temperature values ​​are smoothed using a time-sliding window to eliminate the impact of instantaneous fluctuations. Then, a node spatial adjacency graph is constructed, adjacent temperature gradients are calculated, and pairs of nodes with abnormal gradients are identified. Finally, hotspot stability verification confirms the true hot and cold nodes. This temperature gradient-based hotspot identification method can accurately locate regions of load imbalance, providing clear source and target nodes for subsequent task migration.

[0013] The data dependency topology of the tasks to be scheduled is obtained, and the transmission energy cost and state synchronization latency penalty for task migration between hot and cold nodes are calculated. By analyzing the data dependencies of the tasks, factors such as data transmission volume, network bandwidth, and state reconstruction time during the migration process are evaluated. The energy cost and latency penalty of task migration are comprehensively calculated to provide a quantitative basis for optimizing the migration path. The energy cost reflects the resource consumption of the migration, while the latency penalty reflects the impact on the real-time performance of the task. Together, they determine the benefits and costs of the migration.

[0014] Based on transmission energy consumption costs and state synchronization delay penalties, combined with temperature gradients, a multi-path heat diffusion network for task migration is constructed. This network describes the branching paths of tasks from hot nodes to multiple cold nodes, with each path assigned a conductivity and flow quota calculated based on heat diffusion theory. By analogizing the migration problem to the heat conduction process in a thermal system, this network achieves natural optimization of load balancing. Heat always tends to flow from high-temperature regions to low-temperature regions until the system reaches thermal equilibrium, a physical characteristic highly consistent with the goal of load balancing.

[0015] Based on multi-path heat spread networks and epidemic situation data, a cross-node task decomposition scheme and a spatiotemporally coupled resource reservation strategy are generated. Epidemic situation data is analyzed to predict future computing needs. The target node set is determined according to the heat spread network. Dependency parsing is performed on the task computation graph to identify subtasks that can be parallelized and those that must be serialized. An adaptive task decomposition scheme with computational granularity is generated, and computing resources are reserved simultaneously in both spatial and temporal dimensions to ensure efficient task execution and optimal resource utilization.

[0016] By collecting resource usage pulse sequences from each node during task execution in real time, wavelet transform is used to identify periodic fluctuation patterns and sudden peak characteristics in resource usage. Multi-scale wavelet decomposition is performed on the resource usage data to extract feature information from different frequency bands. Autocorrelation analysis is used to identify periodic patterns, and amplitude analysis is used to capture sudden events, providing real-time feedback for dynamic optimization of resource scheduling. This wavelet analysis-based resource monitoring method can accurately capture the time-frequency characteristics of resource usage and distinguish between normal fluctuations and abnormal peaks.

[0017] Based on the frequency drift of the periodic fluctuation pattern and the amplitude changes of the burst peak characteristics, the task decomposition granularity in the task decomposition scheme and the node load allocation weights in the multi-path heat dissipation network are dynamically adjusted. By analyzing the changing trends of resource occupancy patterns, the task granularity and node load allocation strategies are adaptively adjusted, enabling the scheduling system to flexibly respond to dynamic changes in resource usage patterns. The frequency drift reflects the changes in the system's periodic load, while the burst intensity index reflects the impact of unexpected loads; together, they guide the real-time optimization of the scheduling strategy.

[0018] Based on the adjusted task decomposition granularity and node load allocation weights, the resource scheduling system is controlled to perform task migration and resource allocation, achieving thermal balance and efficient utilization of computing resources. According to the optimized scheduling strategy, the urgency of task migration is calculated, a migration queue is constructed, runtime snapshots are captured, differential compression transmission is executed, migration effects are monitored, and the achievement of thermal balance is verified, completing the closed-loop control process of the entire resource scheduling. Achieving thermal balance signifies that the cluster load distribution is becoming more balanced, maximizing the overall system efficiency and providing stable and efficient computing power support for epidemic prevention and control tasks.

[0019] In this embodiment of the invention, the detailed implementation steps for constructing the load temperature field model of the computing node cluster include: The CPU utilization, memory utilization, disk I / O throughput, and network I / O throughput of each computing node are obtained to form a node load feature vector. The load feature vector is a multi-dimensional data structure describing the resource usage of a node, comprehensively reflecting the working status of the computing node. The acquisition process involves real-time data collection through a system monitoring agent, followed by preprocessing to remove noise and outliers. CPU utilization reflects the processor's workload, memory utilization indicates memory resource usage, and disk I / O and network I / O throughput reflect the intensity of storage and communication resource usage. These four metrics together constitute a complete profile of the node load, providing the raw data foundation for subsequent temperature mapping.

[0020] The node load feature vector is normalized across its dimensions, and weight coefficients are assigned based on the impact of different resource types on system performance. Normalization is a crucial step in eliminating differences in the dimensions of different resource indicators, ensuring data comparability across dimensions. The process employs Min-Max normalization, mapping each indicator value to the [0,1] interval. Weight allocation is based on the importance of resources to system performance and their potential bottleneck status; typically, CPU and memory have higher weights, followed by I / O. The weight allocation also considers resource preferences specific to application types; for example, compute-intensive tasks prioritize CPU, while data-intensive tasks focus more on I / O performance. This weighting scheme ensures that the temperature model accurately reflects the system's true load condition.

[0021] The weighted Euclidean norm of the nodal load feature vector is calculated and denoted as the nodal load intensity value. The load intensity value is a crucial step in compressing multidimensional load characteristics into a single indicator, facilitating subsequent temperature mapping and visualization analysis. The calculation uses a weighted Euclidean norm formula, combining the various dimensional indicators according to their weights into a comprehensive score: ; in, This is the weighted load strength value. For the first A normalized load metric (such as CPU utilization). These are the corresponding weighting coefficients. This calculation method considers both the absolute magnitude of each indicator and the differentiated contribution of different resources to the overall load, resulting in a load intensity value that comprehensively reflects the overall load level of the node.

[0022] This involves mapping node load intensity values ​​to temperature ranges, establishing a nonlinear mapping function between load intensity and temperature values. This nonlinear mapping function has a steeper gradient in high-load regions. Temperature mapping is a core step in thermodynamic modeling, transforming abstract load values ​​into intuitive temperature concepts. The mapping employs an improved sigmoid function, with a steeper gradient designed in high-load regions, making high-load conditions more sensitive to temperature changes and facilitating the timely detection of potential hotspots. The general form of the mapping function is: ; in, The mapped temperature value. This is the load strength value. and These are the lower and upper limits of the temperature, respectively. and Parameters for controlling the curve shape, Determine the steepness of the curve. Determine the location of the midpoint of the curve. By appropriately adjusting the parameters, the mapping function can exhibit a steeper gradient in the high-load region (such as when the load intensity is greater than 0.7), thereby enhancing the monitoring sensitivity of the critical state.

[0023] Temperature values ​​are assigned to corresponding computing nodes, constructing a discrete temperature distribution on the spatial topology of the node cluster. This discrete temperature distribution is a crucial step in the initial formation of the temperature field, organizing individual temperature values ​​into a spatially related distribution structure. The construction process first establishes a spatial model based on the physical or logical topological relationships of the nodes; this could be the physical layout of a data center or the logical structure of network connections. Then, the calculated temperature values ​​are assigned to the corresponding node locations, forming a discrete set of temperature distribution points. This distribution visually represents the load status of each node within the cluster; high-temperature areas represent hotspots with concentrated loads, while low-temperature areas represent resource-rich colds, providing a spatial perspective for subsequent analysis.

[0024] Spatial interpolation is performed on discrete temperature distributions to generate a continuous load temperature field model, which supports temperature value lookup at any spatial location. The continuous temperature field is a smooth extension of the discrete temperature points; interpolation fills in the temperature estimates of unsampled points in space, forming a complete temperature field. The interpolation employs a modified radial basis function (RBF) method, considering the connectivity between nodes and distance attenuation characteristics to generate a smooth temperature field that conforms to thermodynamic laws. The specific formula is as follows: ; in, Spatial location interpolated temperature, Given the known node positions, These are the weighting coefficients. Radial basis functions are typically chosen, such as Gaussian functions or thin-plate spline functions. The continuous temperature field model not only visualizes the load distribution but also supports temperature queries and gradient calculations at arbitrary locations, providing a theoretical basis for hotspot identification and path optimization. This model transforms abstract resource loads into intuitive thermodynamic concepts, enabling the solution of complex resource scheduling problems using thermodynamic principles, thus possessing strong theoretical support and practical value.

[0025] In this embodiment of the invention, the detailed implementation steps for identifying hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold include: The load temperature values ​​of each computing node in the load temperature field model are smoothed using a time-sliding window to eliminate instantaneous load fluctuations, resulting in smoothed node temperature values. Time smoothing is a crucial preparatory step for identifying stable hotspots, filtering out short-term fluctuations through time-dimensional data fusion. The process employs an exponentially weighted moving average (EWMA) algorithm, averaging historical temperature values ​​with decreasing weights, ensuring the most recent data has the highest weight, while the influence of historical data decreases over time. The smoothing formula is: ; in, For a moment The smooth temperature value, This is the original temperature value. The smoothing factor is typically set between 0.1 and 0.3. The window size is dynamically adjusted based on load variation characteristics, usually ranging from 10 to 30 sampling periods. Smoothing effectively eliminates short-term load fluctuations and measurement noise, making the identified hotspots time-stable and avoiding over-response to temporary load peaks.

[0026] A spatial adjacency graph is constructed for the computing node cluster, and the set of neighboring nodes for each node is determined based on this graph. The spatial adjacency graph is the topological foundation for analyzing temperature gradients, describing the spatial relationships and connection structures between nodes. The construction process is based on the physical location of nodes or network topology, defining the adjacency relationships between nodes to form an undirected or directed weighted graph structure. Physical spatial adjacency is typically defined based on Euclidean distance or Manhattan distance, while network topology adjacency is based on direct connections through communication links. For complex network structures, link bandwidth and latency can also be considered as adjacency weights to more accurately reflect the actual relationships between nodes. After the adjacency graph is constructed, each node is associated with a set of neighboring nodes, providing a clear computational object for subsequent temperature gradient calculations.

[0027] The difference between the smoothed temperature values ​​of each node and all nodes in its neighboring node set is calculated and denoted as the adjacency temperature gradient. Temperature gradient calculation is a core step in hotspot identification, quantifying the degree of load imbalance between adjacent nodes. The calculation process involves calculating the difference in smoothed temperature values ​​for each pair of adjacent nodes (i,j), forming an adjacency temperature gradient matrix. For weighted adjacency graphs, the gradient values ​​can be combined with link weights to reflect the actual migration difficulty. Gradient calculation considers not only absolute temperature differences but also the standardization of spatial distance, yielding the rate of temperature change per unit distance, which better aligns with the physical definition of a gradient in thermodynamics. These temperature gradient values ​​directly reflect the degree of load unevenness, providing a quantitative basis for identifying abnormal hotspots and cold spots.

[0028] Pairs of adjacent nodes whose absolute temperature gradient values ​​exceed a preset threshold are designated as candidate temperature gradient anomaly pairs. Gradient threshold screening is a fast method for initially identifying anomalous hotspots, quickly pinpointing significantly unbalanced regions by setting a reasonable threshold. The screening process compares the absolute value of the temperature gradient of each pair of adjacent nodes with the preset threshold; when the gradient exceeds the threshold, the pair is marked as a candidate anomaly pair. The threshold setting is dynamically adjusted based on system characteristics and load distribution patterns, typically choosing the upper quartile of the temperature gradient distribution or the mean plus two standard deviations. To improve the robustness of the screening, consistency verification of the gradient direction can be added to ensure that anomalous gradients are not caused by random fluctuations. Candidate anomaly pairs provide preliminary screening results for subsequent hotspot verification, significantly reducing computational complexity.

[0029] For nodes with higher temperature values ​​among candidate temperature gradient anomalies, load persistence verification is performed. The percentage of time the node's temperature remains high within a preset verification period is calculated and denoted as hotspot stability. Hotspot stability verification is a crucial step in confirming genuine hotspots, distinguishing between temporary load peaks and persistent hotspots through time-dimensional persistence analysis. For each candidate hotspot node, the verification process analyzes its historical temperature value sequence and calculates the percentage of time the temperature exceeds the hotspot threshold. The verification duration is set according to application characteristics, typically ranging from a few minutes to tens of minutes. Hotspot stability calculation considers not only the proportion of time exceeding the threshold but also the average temperature value, comprehensively assessing the severity and persistence of the hotspot. This spatiotemporal dual verification mechanism significantly improves the accuracy and reliability of hotspot identification, avoiding overreaction to temporary fluctuations.

[0030] Nodes with a stability greater than a preset stability threshold are marked as hotspot nodes, and their corresponding nodes with lower temperature values ​​are marked as coldspot nodes. The final marking is the output of the hotspot identification conclusion, determining the final hotspot-coldspot node pairs based on the stability threshold. The marking process first filters stable hotspots based on their stability, then selects coldspots that meet certain conditions from the adjacent low-temperature nodes. Coldspot selection considers factors such as the absolute value of temperature difference, the resource margin of the coldspot, and connection bandwidth to ensure that the selected coldspots have sufficient resources to receive tasks migrated from the hotspots. The final marked node pairs constitute the basic migration unit for load balancing. Hotspots serve as the source of task migration, and coldspots serve as the destination of task migration; together, they define the direction and scope of load migration, providing clear execution targets for subsequent task migration strategies.

[0031] In this embodiment of the invention, the detailed implementation steps for the transmission energy consumption cost and state synchronization delay penalty of the computing task migrating between hot and cold nodes include: The data dependency topology graph of the task to be scheduled is analyzed to extract the required input dataset size and intermediate state data size. Data dependency analysis is a fundamental step in assessing migration costs, clarifying data transmission requirements through topology graph analysis. The analysis process uses a graph traversal algorithm to process the computation graph structure of the task, identify data dependencies between nodes, and quantify the data size of each dependent link. The input dataset size reflects the amount of data that needs to be transmitted at the initial start of the task, while the intermediate state data size reflects the memory state size during task execution; both together determine the data transmission requirements during migration. For complex AI computation graphs, critical path analysis and data flow estimation are also employed to more accurately predict data interaction patterns in actual operation. This data size information provides a quantitative basis for subsequent energy consumption and latency calculations.

[0032] Obtain network bandwidth and latency parameters between hot and cold nodes. Network parameter acquisition is a crucial step in evaluating transmission efficiency, obtaining real-world network performance data through measurement or querying. The acquisition process combines active measurement and historical statistics, considering both static network topology and dynamic performance changes. Bandwidth parameters reflect the maximum data transmission rate, while latency parameters reflect network response time; both together determine the time overhead of data transmission. For complex network environments, bandwidth volatility and sharing characteristics are also considered, using statistical models to predict actual available bandwidth, improving the accuracy of the evaluation. These network parameters provide the fundamental conditions for calculating energy costs and latency penalties, reflecting the physical constraints of migration.

[0033] Based on the input dataset size, network bandwidth, and power consumption per unit data transmission, the data transmission energy consumption during the task migration process is calculated and denoted as the basic transmission energy consumption. Calculating the basic transmission energy consumption is a crucial step in evaluating migration energy costs, directly reflecting the resource consumption of data movement. The calculation employs a product model of data volume, transmission rate, and unit energy consumption, considering the energy consumption characteristics of network devices during transmission. ; in, Based on transmission energy consumption, The size of the input dataset (in bytes). Power consumption per unit of data transmission (watts / byte / second). This represents network bandwidth (bytes per second). Power consumption per unit data transmission is determined based on network device type and operating efficiency, typically estimated using measured data or energy efficiency models. This calculation method considers the dual impact of data volume and transmission efficiency on energy consumption, accurately reflecting the differences in energy consumption across various migration scenarios.

[0034] Based on the size of the intermediate state data and the memory write rate of the cold node, the reconstruction time cost of the state data at the cold node is calculated and denoted as the state reconstruction latency. The state reconstruction latency calculation is a crucial step in evaluating the migration time cost, reflecting the preparation time required for task resumption. The calculation considers the size of the state data and the processing capacity of the target node, employing a simple ratio model: ; in, For state reconstruction delay, The size of the intermediate state data (in bytes). The memory write rate (bytes / second) for cold nodes is given. For complex state data structures, data dependencies and the possibility of parallel reconstruction are also considered, and latency estimation is optimized through critical path analysis. State reconstruction latency directly affects the recovery time after task migration and is an important indicator for evaluating the impact of migration on real-time performance.

[0035] The sum of the base transmission energy consumption and the energy consumption corresponding to the state reconstruction delay is used as the transmission energy cost. Total energy consumption calculation is the final step in comprehensively evaluating energy expenditure, integrating the energy consumption of both the transmission and reconstruction phases. The calculation converts the state reconstruction delay into its corresponding energy consumption, taking into account the additional power consumption of the target node during the reconstruction process. ; in, For transmission energy consumption costs, This represents the additional power consumption (watts) during the reconstruction process. Additional power consumption considers the energy overhead of operations such as memory access, CPU processing, and cache synchronization, and is typically determined based on node-based benchmark power consumption tests. This comprehensive calculation method fully reflects the energy cost of task migration, providing quantitative guidance for energy efficiency optimization.

[0036] The state synchronization delay penalty is calculated by weighting the network latency and state reconstruction latency together with the task's real-time requirements. Delay penalty calculation is the final step in evaluating the impact of migration on task performance, using a weighted model to reflect the latency sensitivity of different scenarios. The calculation combines network transmission latency and state reconstruction latency into a total latency, and determines the penalty weights based on the task's real-time requirements. ; in, As a penalty for state synchronization delay, Due to network latency, The weighting coefficients for reconstruction delay, The real-time weighting of tasks is typically determined based on the task's Service Level Agreement (SLA). Tasks with high real-time requirements are assigned higher penalty weights, causing the scheduling system to evaluate migration decisions more carefully. This penalty mechanism, which considers task characteristics, ensures that the migration strategy balances system efficiency and task performance, providing a reasonable cost assessment for subsequent hot-spreading network construction.

[0037] In this embodiment of the invention, the detailed implementation steps for constructing a multipath heat dissipation network for task migration include: A heat diffusion topology graph is constructed, using hot nodes as heat sources and cold nodes as cold sinks. This topology graph construction is a fundamental step in network modeling, mapping the load balancing problem to a heat conduction problem through physical analogy. The construction process first identifies hot nodes as heat sources and cold nodes as cold sinks, then establishes the topology based on the network connections between nodes. The topology graph is represented as a directed weighted graph, where nodes correspond to computational resource points, edges correspond to possible migration paths, and edge weights reflect path characteristics. For complex network environments, indirect connections and multi-hop paths are also considered, constructing a complete diffusion probability graph. This thermodynamic-based topology mapping intuitively transforms the load balancing problem into a physical process of heat flowing from high to low temperatures, providing a theoretical foundation for subsequent path optimization.

[0038] In the heat diffusion topology, a heat diffusion resistance coefficient is calculated for each path from the heat source to the cold trap. This coefficient is determined by a normalized weighted average of transmission energy cost and state synchronization delay penalty. Calculating the heat diffusion resistance is a core step in path evaluation, quantifying the difficulty of task migration. The calculation process first normalizes the transmission energy cost and delay penalty to make these two heterogeneous indicators comparable. Then, weights are assigned according to scheduling strategy preferences, emphasizing different indicators for scenarios prioritizing energy efficiency and real-time performance. Finally, the weighted sum is calculated as the heat diffusion resistance coefficient for each path. The resistance coefficient calculation considers the matching degree between task characteristics and system state; high-energy-consumption or high-latency paths are assigned higher resistance, reducing the likelihood of heat diffusion through these paths. This resistance calculation method based on multi-objective optimization allows the heat diffusion model to flexibly adapt to different scheduling requirements.

[0039] Based on the temperature gradient and the thermal diffusion drag coefficient, the heat transfer rate of each path is calculated. This heat transfer rate is directly proportional to the temperature gradient and inversely proportional to the thermal diffusion drag coefficient. Calculating the heat transfer rate is a crucial step in path capacity assessment, directly determining the flow distribution for load migration. The calculation is based on a physical analogy of Fourier's law of thermal conduction, combining the driving force of the temperature gradient and path drag to quantify the heat transfer capacity of each path. ; in, For path thermal conductivity, This is the proportionality coefficient. The temperature gradient at both ends of the path, This represents the thermal diffusion drag coefficient of the path. The temperature gradient reflects the driving force of load imbalance, while the drag coefficient reflects the difficulty of migration; the ratio of the two determines the actual conduction capacity of the path. This calculation method based on a physical model naturally balances the system equilibrium requirements and migration costs, making load allocation more consistent with the thermodynamic optimality principle.

[0040] Based on thermal conductivity, a task migration flow quota is allocated to each path. This quota represents the proportion of task load that path can handle. Quota allocation is the final step in path capacity planning, transforming the abstract conductivity into a concrete task allocation ratio. The allocation process first normalizes the thermal conductivity of each path so that the sum is 1; then, based on the normalized value and the total migration load, the specific quota for each path is calculated; finally, considering the actual capacity limitations of the paths, necessary adjustments are made to ensure the rationality of the allocation. For heterogeneous tasks, task divisibility and minimum granularity constraints are also considered to ensure that the allocation results are executable in the actual system. This conductivity-based adaptive allocation mechanism allows efficient paths to receive more task flow, naturally realizing the optimal path-first allocation strategy.

[0041] By integrating the heat spread topology map with the task migration traffic quotas for each path, a multi-path heat spread network is generated. This network includes the distribution paths and traffic allocation schemes for tasks moving from hot nodes to multiple cold nodes. The heat spread network integration is the final step in model building, forming a complete load balancing strategy. The integration process combines the topology, path characteristics, and traffic quotas into a unified network model, preserving the path connections while including specific allocation strategies. For complex multi-source and multi-sink scenarios, network flow algorithms are used to optimize the overall allocation, ensuring global optimality. The resulting multi-path heat spread network is a complete scheduling strategy model that intuitively describes how tasks are distributed from hot nodes to multiple cold nodes, and the specific load proportion borne by each path. This multi-path distribution strategy significantly improves the system's load balancing efficiency and robustness, avoiding the bottleneck problem of single-point transmission.

[0042] In this embodiment of the invention, the detailed implementation steps for generating a cross-node task decomposition scheme and a spatiotemporally coupled resource reservation strategy include: Based on the epidemic situation data, the growth curve of computing task demand in various regions is predicted within the future time window. Demand forecasting is a forward-looking step in resource planning, using epidemic data analysis to anticipate changes in computing demand. The forecasting process first analyzes the correlation between historical epidemic data and computing load, establishing a mapping relationship between epidemic indicators and computing demand; then, it combines infectious disease dynamics models to predict the development trend of the epidemic, further deriving the future curve of computing demand changes; finally, it considers the impact of regional differences and intervention measures to form a refined forecast for each region. The forecasting model combines time series analysis and machine learning methods, capturing both the periodicity and trend of demand while considering the nonlinear impact of unforeseen events. This demand forecasting based on the epidemic situation provides a time-dimensional planning basis for resource reservation, ensuring that the system can respond to load changes in advance.

[0043] Based on the task migration traffic quota in the multi-path heat spread network, the target cold node set and the task allocation load of each node are determined. Target node planning is a fundamental step in spatial resource allocation, clarifying the specific objectives and proportions of task migration. The planning process is based on the traffic quota of the heat spread network, combined with the actual resource surplus of cold node, to determine the final target node set and the load allocation load of each node. The planning considers the differences in computing power of nodes, current load levels, and future demand forecasts to ensure that allocation does not lead to the creation of new hotspots. For important tasks, a redundant allocation strategy is also adopted to improve reliability through appropriate replication. This node planning method based on multi-objective optimization balances system efficiency and task reliability, providing clear objective guidance for task decomposition.

[0044] Dependency analysis is performed on the computation graph of the task to be migrated, identifying clusters of subtasks that can be executed in parallel and critical paths that must be executed sequentially. Task dependency analysis is a key step in optimizing execution efficiency, uncovering the parallel potential of tasks through graph structure analysis. The analysis process first represents the task as a directed acyclic graph, with nodes as computational units and edges as data dependencies; then, topological sorting is performed to identify execution levels and parallelization possibilities; finally, critical path analysis identifies the sequential task chains that determine the total execution time. For complex AI models, algorithm characteristic analysis is also combined to identify internal data parallelism and model parallelism opportunities. This fine-grained task dependency analysis provides structural guidance for efficient distributed execution, enabling task decomposition to maximize the utilization of parallel resources and reduce unnecessary waiting.

[0045] Based on the computing power and task allocation load of each node in the target cold node set, the subtask cluster is decomposed into multiple subtask units with adaptive computational granularity, generating a task decomposition scheme. Task decomposition is the core step in execution optimization, improving computational efficiency through reasonable granularity partitioning. The decomposition process first evaluates the computational complexity and resource requirements of each subtask, and then determines the optimal decomposition granularity based on the processing power and allocated load of the target nodes. For subtasks with different characteristics, differentiated decomposition strategies are adopted: computationally intensive tasks tend to have finer granularity, while communication-intensive tasks maintain larger granularity to reduce communication overhead. The decomposition also considers the principle of data locality, minimizing cross-node data dependencies and reducing communication costs. This computationally granularity-adaptive decomposition method enables tasks to better adapt to heterogeneous computing environments, improving overall execution efficiency.

[0046] Based on the growth curve of computing task demand and the estimated execution time of the task decomposition scheme, resources are reserved in the target cold node set. This resource reservation simultaneously locks the computing cores, memory capacity, and network bandwidth of the nodes, and marks the start time and duration of the reserved resources, forming a spatiotemporally coupled resource reservation strategy. Resource reservation is a key step in ensuring execution, ensuring smooth task execution by locking resources in advance. The reservation process is planned in two dimensions: time and space, based on task execution time estimation and demand growth prediction. Spatially, it locks multiple types of resources on different nodes, and temporally, it marks the start and end times of resource occupancy. The reservation strategy adopts a sliding time window mechanism, dynamically adjusting the reservation based on execution progress and real-time monitoring, ensuring resource availability while avoiding resource waste caused by over-reservation. For high-priority urgent tasks, a resource preemption mechanism is also set up to ensure timely response in case of emergencies. This spatiotemporally coupled reservation strategy significantly improves the accuracy and flexibility of resource allocation, providing a stable resource environment for task execution.

[0047] In this embodiment of the invention, the detailed implementation steps for identifying the periodic fluctuation patterns and sudden peak characteristics of resource usage include: The CPU utilization, memory utilization, and network throughput of each node are collected in real time at a preset sampling frequency, forming a multi-dimensional resource utilization pulse sequence. Resource monitoring is a fundamental step in performance analysis, capturing resource usage patterns through high-precision sampling. The collection process employs a distributed monitoring agent program, synchronously collecting multiple resource indicators at a preset frequency (typically 1-10 seconds / time), forming a time-aligned multi-dimensional data sequence. For critical nodes or periods of high volatility, the sampling frequency can be dynamically increased to capture finer-grained changes. The collected data undergoes preprocessing to eliminate outliers and sampling errors, ensuring data quality. The multi-dimensional pulse sequence not only includes the usage of individual resource types but also retains correlation information between different resources, providing a comprehensive data foundation for subsequent pattern analysis.

[0048] Wavelet transform is performed on each dimension of the multidimensional resource occupancy pulse sequence to obtain multi-scale wavelet decomposition coefficients. Wavelet analysis is the core step in time-frequency characteristic extraction, revealing the scale characteristics of the signal through multi-resolution decomposition. The decomposition process selects suitable wavelet basis functions (such as Daubechies wavelets or Symlet wavelets) and performs multi-level wavelet transforms on the resource occupancy sequence, decomposing the original signal into sub-signals with different frequency ranges. The number of decomposition levels is dynamically adjusted according to the sequence length and analysis objectives, typically 3-5 levels. Multi-scale wavelet decomposition can simultaneously capture both local details and global trends of the signal, making it particularly effective for complex signals like resource occupancy that exhibit both periodicity and suddenness. The resulting coefficient matrix contains rich time-frequency information, providing a mathematical foundation for pattern recognition.

[0049] In multi-scale wavelet decomposition coefficients, low-frequency components are extracted as trend terms for resource occupancy, and mid-frequency components are extracted as candidates for periodic fluctuations. Feature separation is a key step in pattern recognition, distinguishing different types of resource changes through frequency band division. The separation process is based on the natural frequency band division of wavelet decomposition: low-frequency components (approximate coefficients) correspond to long-term trends, mid-frequency components (detail coefficients at intermediate levels) correspond to periodic fluctuations, and high-frequency components (detail coefficients at the smallest scale) correspond to instantaneous changes. The trend term reflects the basic level and long-term direction of resource use, while the candidates for periodic fluctuations contain potential regular change patterns. Separating these two facilitates targeted analysis. This frequency-based pattern separation method effectively improves the accuracy and interpretability of feature extraction.

[0050] Autocorrelation analysis is performed on candidate periodic fluctuations to calculate the peak positions of the autocorrelation function. The time intervals corresponding to these peak positions are recorded as the characteristic period of the periodic fluctuation. Period detection is the core step in pattern identification, revealing repetitive patterns through the correlation between the signal and itself. The detection process calculates the autocorrelation function for the intermediate frequency component sequence, identifies significant peaks in the function, and these peaks correspond to repetitive patterns in the sequence. The characteristic period is determined by the position of the first significant peak, representing the basic repetitive period of resource occupancy. For multi-period signals, spectral analysis is used for further decomposition to identify nested periods at different time scales. Period detection not only identifies the period length but also assesses period stability and signal-to-noise ratio, providing a quantitative evaluation of period significance. This autocorrelation-based period analysis method can effectively identify regular patterns in resource use, providing important references for prediction and optimization.

[0051] Mid-frequency components with fluctuation amplitudes exceeding a preset periodic fluctuation threshold are marked as periodic fluctuation patterns. Pattern confirmation is the final step in feature extraction, determining valid periodic patterns through threshold screening. The confirmation process calculates the fluctuation amplitude corresponding to the feature period and compares it with a preset threshold. Only when the fluctuation significantly exceeds the background level is it marked as a valid periodic pattern. Threshold settings are based on signal noise levels and application requirements, typically determined adaptively using statistical methods. For multidimensional resource data, phase relationships between different dimensions are also analyzed to identify collaborative fluctuation patterns among resources. The finally confirmed periodic fluctuation patterns include period length, fluctuation amplitude, and phase information, comprehensively describing the regularity of resource use and providing a basis for periodic prediction for dynamic adjustments.

[0052] In the high-frequency components of multi-scale wavelet decomposition coefficients, abrupt amplitude changes are identified, and points with abrupt amplitudes exceeding a preset threshold are marked as sudden peak features. Abrupt change detection is the core step in anomaly identification, discovering unusual events through drastic signal changes. The detection process analyzes the amplitude distribution of high-frequency components to identify abrupt changes that significantly deviate from normal levels. A sliding window local comparison method is used to calculate the deviation between the current point and the window's statistical characteristics; when the deviation exceeds a preset threshold, it is marked as an abrupt change point. To improve the robustness of detection, multi-scale analysis is combined to verify the consistency of abrupt changes at different time scales, reducing false alarms. The feature description of abrupt change points includes the abrupt change time, amplitude, duration, and contextual information, comprehensively characterizing unusual changes in resource use. This wavelet-based abrupt change detection method exhibits high sensitivity to sudden events in resource use, providing timely early warnings for dynamic responses.

[0053] In this embodiment of the invention, the detailed implementation steps for dynamically adjusting the task decomposition granularity and the node load allocation weights in the multipath heat dissipation network include: The difference between the current characteristic period and the initial characteristic period of a periodic fluctuation pattern is calculated and denoted as the frequency drift. Frequency tracking is a fundamental step in change detection, revealing the evolutionary trend of resource use patterns through period comparison. The calculation process compares the currently observed characteristic period with the characteristic period of the system's initial state or the previous period to determine the direction and magnitude of the periodic change. For complex periodic systems, phase-locked loop (PLL) technology is used to continuously track periodic changes, improving sensitivity to gradual changes. The sign of the frequency drift reflects whether the period is getting longer or shorter, while the absolute value reflects the degree of change. This dynamic periodic tracking method can promptly capture changing trends in resource use patterns, providing directional guidance for task granularity adjustments.

[0054] The frequency and average amplitude of sudden peak characteristics within a preset statistical window are recorded as the suddenness intensity index. Suddenness analysis is the core step in anomaly assessment, quantifying the unconventional nature of resource fluctuations through statistical characteristics. The analysis process, within a sliding time window, statistically analyzes the density and intensity characteristics of abrupt change points to form a comprehensive suddenness index. The window size is set based on system response time and resource characteristics, typically ranging from several minutes to tens of minutes. The suddenness intensity index considers both the frequency of sudden events (reflecting instability) and the amplitude of sudden events (reflecting the degree of impact), comprehensively assessing the system's fluctuation state. For multi-resource collaborative analysis, the temporal correlation of sudden events from different resources is also considered to identify cascading effects. This multi-dimensional suddenness assessment method provides a quantitative measure of system instability, guiding the dynamic adjustment of scheduling strategies.

[0055] Based on the direction and magnitude of frequency drift, the direction of task decomposition granularity adjustment is determined: when the frequency drift is positive, the task decomposition granularity is increased; when the frequency drift is negative, the task decomposition granularity is decreased. Granularity adjustment is a key step in execution optimization, improving computational efficiency by adapting to resource rhythm. The adjustment follows the principle of matching resource cycle with task granularity: when the resource occupancy cycle becomes longer (positive frequency drift), the task granularity is increased to reduce scheduling overhead; when the cycle becomes shorter (negative frequency drift), the task granularity is decreased to improve response flexibility. The adjustment magnitude is proportional to the frequency change; significant changes lead to more aggressive adjustments, while minor changes lead to more conservative adjustments. This cycle-adaptive granularity adjustment strategy allows the task execution rhythm to naturally synchronize with changes in resource availability, reducing resource contention and waiting time, and improving overall execution efficiency.

[0056] Based on the burst intensity index, an adjustment coefficient for node load allocation weights is calculated, which is inversely proportional to the burst intensity index. Weight adjustment is the core step in allocation optimization, improving system stability by responding to burst characteristics. The calculation employs an inverse proportionality model, ensuring that nodes with high burst intensity receive lower allocation weights, reducing their task load. The adjustment coefficient calculation considers both the frequency and intensity of bursts, imposing stricter constraints on nodes with frequent and significant fluctuations. Simultaneously, a smoothing mechanism is incorporated into the adjustment process to avoid frequent adjustments caused by over-responding to temporary fluctuations. This adaptive adjustment strategy based on burst characteristics enables the system to proactively avoid unstable nodes, allocating tasks to more reliable resources and improving overall execution stability and predictability.

[0057] The task decomposition granularity adjustment direction and adjustment coefficients are applied to the task decomposition scheme to update the computational granularity of sub-task units. Scheme update is a key step in optimization implementation, translating adjustment decisions into concrete execution plans. The update process first modifies the granularity parameters of the task decomposition based on the granularity adjustment direction and magnitude; then, it re-divides sub-task boundaries, merging or splitting existing task units; finally, it updates the dependencies between tasks to ensure logical correctness. The update considers the minimum divisible granularity of tasks and data dependency constraints to ensure the adjustment results are executable. For tasks currently being executed, an incremental adjustment strategy is adopted, modifying only the parts that have not yet started to reduce the impact of interruptions. This gradual scheme optimization method allows the system to gradually adjust to a better task decomposition structure without affecting current execution.

[0058] Based on the updated computational granularity and the current load temperature of each node, the task migration traffic quota for each path in the multi-path heat diffusion network is recalculated, and the node load allocation weights are updated. Weight updates are the final step in allocation optimization, comprehensively considering granularity changes and node states to rebalance task allocation. The update process first reassesses the system's thermal distribution based on the latest node load temperature values; then, combining the updated task granularity and node burst characteristics, the conductivity and traffic quotas in the heat diffusion network are recalculated; finally, new load allocation weights are generated to guide subsequent task allocation. The update employs a gradual strategy to avoid system oscillations caused by large adjustments, gradually approaching the optimal state through small iterations. This continuous optimization and dynamic adjustment mechanism enables the system to flexibly respond to changing resource conditions and task characteristics, always maintaining a near-optimal operating state.

[0059] In this embodiment of the invention, the detailed implementation steps for identifying clusters of subtasks that can be executed in parallel and critical paths that must be executed serially in the computation graph include: Representing a computational graph as a directed acyclic graph (DAG) involves nodes representing computational task units and edges representing data dependencies. Graph structure modeling is a fundamental step in task analysis, capturing the logical structure of the task through formal representation. The modeling process first identifies the basic units of the computational task, which can be functions, modules, or algorithm components; then it analyzes the data flow and control dependencies between units, establishing directed edge connections; finally, it verifies the acyclicity of the graph to ensure execution feasibility. For complex AI tasks, modeling also considers performance characteristics such as operator computational complexity, memory usage, and data transfer volume, providing richer information for subsequent analysis. As a standard representation of computational tasks, the directed acyclic graph intuitively describes the execution order and dependency constraints of the task, providing a structured foundation for parallelism analysis.

[0060] A topological sort is performed on a directed acyclic graph (DAG) to obtain the execution hierarchy sequence of task units. Hierarchical identification is a key step in parallel analysis, revealing the natural execution stages of tasks through sorting. The sorting employs a classic topological sorting algorithm, assigning task units to different execution levels based on node in-degree and dependencies. Tasks within each level are independent and can be executed in parallel; dependencies exist between different levels, requiring sequential execution. The sorting result forms a hierarchical execution plan, clearly showing the parallel potential and execution constraints of the tasks. For complex graph structures, the sorting also considers a critical path priority strategy, prioritizing tasks with a greater impact on total execution time to improve overall efficiency. This topology-based hierarchical analysis method provides a systematic analytical framework for identifying parallel opportunities.

[0061] Within the same execution level, sets of task units without data dependencies are identified and marked as clusters of subtasks that can be executed in parallel. Parallel set identification is the core step in optimization decomposition, discovering independent task groups through dependency analysis. The identification process, within each execution level, analyzes data access patterns and dependencies between task units, grouping tasks without direct or indirect dependencies into the same subtask cluster. For tasks with partial dependencies, further subdivision is performed based on the strength and type of the dependency, forming an optimal parallel organization structure. The division of subtask clusters also considers the principles of load balancing and communication minimization, ensuring that parallel execution fully utilizes resources while reducing coordination overhead. This fine-grained parallelism identification method maximizes the parallel execution potential of tasks, providing structural guidance for efficient distributed processing.

[0062] From the entry node to the exit node of a directed acyclic graph, the cumulative execution time of all paths is calculated, and the path with the longest cumulative execution time is selected as the critical path. Critical path identification is a core step in performance optimization, identifying the task chain that determines the total execution time through time analysis. The identification process first estimates the execution time of each task unit based on historical data or performance models; then, it calculates the cumulative time of all possible paths in the graph and identifies the path with the longest execution time; finally, this path is marked as the critical path, serving as the primary target for optimization. Critical path analysis not only identifies the paths themselves but also calculates the time margin of each non-critical path, providing priority guidance for resource allocation. For dynamic execution environments, probabilistic critical path analysis is also employed to consider the randomness of execution time and more comprehensively assess the criticality of paths. This time-based path analysis method allows optimization to focus on the key links that truly affect overall performance.

[0063] Fine-grained analysis is performed on task units along the critical path to determine if they possess decomposable data-parallel structures. Internal parallelism analysis is a core step in optimizing the critical path, uncovering hidden parallel opportunities through fine-grained decomposition. The analysis delves into the internal structure of task units, studying their algorithmic characteristics and data access patterns to determine if parallelizable computational patterns exist. Typical data-parallel structures include independent loop iterations, divisible data batches, and naturally distributed problem domains, which allow atomic tasks to be further decomposed into parallel subtasks. For numerical computation and machine learning tasks, the analysis also considers the algorithm's numerical stability and parallel speedup to ensure that decomposition does not affect the accuracy of the results. This deep parallelism mining method can overcome the limitations of surface dependencies, discover more parallel execution opportunities, especially for critical path optimization, directly improving overall execution efficiency.

[0064] The process decomposes data-parallel task units along the critical path, generating new sub-task nodes and updating the dependencies in the directed acyclic graph (DAG). Graph structure optimization is the final step in parallel implementation, achieving a more efficient execution mode by reconstructing the computation graph. The optimization process first decomposes the identified parallelizable task units, creating multiple sub-task nodes; then, it re-establishes the dependencies between these sub-tasks and the original nodes, updating the DAG structure; finally, it verifies the correctness and execution efficiency of the updated graph to ensure that the optimization truly delivers performance improvements. The decomposition considers the minimum granularity of data dependencies and the balance between communication overhead, avoiding coordination burdens caused by over-decomposition. For critical path tasks, the decomposition also pays special attention to load balancing and robustness, ensuring that no new performance bottlenecks are created after decomposition. This critical path-based graph optimization method specifically improves the parallelism of the system, reduces the overall execution time, and provides a better task structure for resource scheduling.

[0065] In this embodiment of the invention, the detailed implementation steps for controlling the resource scheduling system to perform task migration and resource allocation include: Based on node load allocation weights, the migration urgency of each subtask unit is calculated. This urgency is determined by the product of the current temperature value of the hot node where the subtask resides and the load contribution of that subtask. Urgency calculation is a quantification step in migration priority, determining the migration order of tasks through comprehensive evaluation. The calculation considers two key factors: the node's temperature value reflects the overall load level, and the subtask's load contribution reflects the impact of a specific task. Their product intuitively represents the potential benefits of task migration. For complex tasks, the calculation also considers migration difficulty and dependency constraints, providing a more comprehensive assessment of migration value. The urgency score provides a reasonable ranking basis for migration queue construction, ensuring that the system prioritizes the most impactful hot tasks and maximizes load balancing effectiveness.

[0066] A task migration queue is constructed based on migration urgency from high to low. Resource reservation time conflicts are detected for target cold nodes of adjacent subtasks in the migration queue. The migration start time is fine-tuned to stagger the resource occupation time periods of conflicting task pairs, generating a conflict-free migration sequence table. Queue optimization is a crucial step in scheduling execution, avoiding resource conflicts through time coordination. The optimization process first constructs an initial queue based on urgency, then detects resource reservation conflicts on target nodes. When multiple tasks compete for the same resource in the same time period, the conflict is resolved by adjusting the start time. Conflict resolution employs a combination of greedy algorithms and dynamic programming to minimize overall latency while prioritizing high-urgency tasks. For tasks with complex dependencies, optimization also considers the overall execution efficiency of the task chain, maintaining the continuity of critical links. This time-coordinated scheduling method significantly reduces waiting time caused by resource contention, improving the smoothness and efficiency of the migration process.

[0067] A runtime state snapshot is captured for subtasks to be migrated on hot nodes. This snapshot includes memory images, register states, and computation progress markers. State capture is a fundamental step in task migration, ensuring computational continuity through snapshot preservation. The capture process pauses computation when the task reaches a safe point, saving complete execution state information, including memory data, register values, and computation progress markers. For long-running tasks, incremental snapshot technology is used, recording only the changed state parts to reduce capture overhead. The snapshot mechanism incorporates fine-grained consistency checkpoints to ensure state integrity and consistency, supporting reliable execution recovery. For memory-intensive tasks, memory page tracking and copy-on-write techniques are also employed to optimize snapshot space efficiency. This comprehensive state preservation method provides the technical foundation for seamless task migration between different nodes, guaranteeing the correctness and consistency of computation results.

[0068] The runtime snapshot and the input dataset are differentially compressed to generate incremental state data packets, which are then transmitted to the target cold node. Simultaneously, a backup copy is maintained on the hot node, with the backup retention period set to the stable operation verification cycle. Data optimization is a key step in improving transmission efficiency, reducing data volume through compression and differential techniques. Optimization first analyzes the overlap between the state data and the original input to avoid duplicate transmission; then, suitable compression algorithms are applied to the differing parts, such as lossless compression for critical data and lossy compression for fault-tolerant data; finally, structured incremental data packets are generated, containing metadata indexes and compressed content. Transmission employs multi-path parallel channels and a breakpoint resumption mechanism to improve transmission reliability and efficiency. The backup strategy retains the original copy on the hot node as a rollback guarantee in case of migration failure; the retention period is based on an empirically set verification cycle, typically ranging from a few minutes to tens of minutes. This optimized data transmission method significantly reduces network overhead and time costs during migration, improving overall migration efficiency.

[0069] Real-time monitoring of the load temperature change rate of each cold node after migration triggers a task rollback mechanism when the change rate exceeds a preset safety threshold. Anomaly monitoring is a crucial step in ensuring execution, identifying potential problems through real-time tracking. The monitoring process continuously collects load metrics from target nodes and calculates the temperature change rate. Abnormally rapid or directional changes are identified as potential risks. Monitoring employs multi-dimensional metric combinations and trend analysis to improve the accuracy and foresight of judgments. The rollback mechanism incorporates a multi-level response strategy, ranging from minor adjustments to complete rollback, selecting appropriate measures based on the severity of the anomaly. For distributed tasks, rollback also considers inter-task dependencies to ensure that the rollback operation does not affect other normally executing tasks. This proactive monitoring mechanism provides real-time assurance for the migration process, preventing system instability or performance degradation caused by improper migration.

[0070] After task migration, the temperature variance of the load temperature field model is calculated and compared with the initial temperature variance before migration. When the reduction in temperature variance reaches a preset thermal balance threshold, thermal balance of computing resources is confirmed. Effect evaluation is the final step in the scheduling closed loop, verifying the optimization results through objective metrics. The evaluation uses temperature variance as the main indicator, intuitively reflecting the degree of load distribution balance; a smaller variance indicates a more balanced system. The calculation and comparison of variance changes before and after migration focuses not only on the absolute value but also on the relative improvement percentage, which better reflects the actual optimization effect. The thermal balance threshold is set based on system characteristics and application requirements, typically representing a 15%-30% reduction in variance. For critical applications, the evaluation also considers comprehensive indicators such as task execution time and resource utilization to comprehensively evaluate the scheduling effect. This data-driven effect verification method provides an objective evaluation basis for the scheduling strategy, ensuring that the system truly achieves the expected optimization goals.

[0071] The above describes the big data AI resource scheduling method for infectious disease prevention and control in the embodiments of this application. The following describes the big data AI resource scheduling system for infectious disease prevention and control in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the big data AI resource scheduling system for infectious disease prevention and control in this application includes: The data acquisition module is used to acquire real-time load data and infectious disease epidemic situation data of each computing node; The temperature field modeling module is used to construct a load temperature field model of the computing node cluster based on real-time load data. The hot spot identification module is used to identify hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold by calculating the temperature gradient between adjacent nodes in the load temperature field model. The task analysis module is used to obtain the data dependency topology graph of the task to be scheduled, and calculate the transmission energy consumption cost and state synchronization latency penalty of the task migrating between hot nodes and cold nodes based on the data dependency topology graph. The heat dissipation network construction module is used to construct a multi-path heat dissipation network for task migration based on transmission energy consumption cost and state synchronization delay penalty, combined with temperature gradient. The task decomposition module is used to generate cross-node task decomposition schemes and spatiotemporal coupled resource reservation strategies based on multi-path heat spread network and epidemic situation data. The spatiotemporal coupled resource reservation strategy locks the computing resources of cold nodes in both time and space dimensions. The resource monitoring module is used to collect the resource usage pulse sequence of each node in real time during task execution, perform wavelet transform on the resource usage pulse sequence, and identify the periodic fluctuation pattern and sudden peak characteristics of resource usage. The dynamic adjustment module is used to dynamically adjust the task decomposition granularity in the task decomposition scheme and the node load allocation weight in the multi-path heat diffusion network based on the frequency drift of the periodic fluctuation pattern and the amplitude changes of the sudden peak characteristics. The execution control module is used to control the resource scheduling system to perform task migration and resource allocation based on the adjusted task decomposition granularity and node load allocation weight, so as to achieve thermal balance and efficient utilization of computing resources. The modules are connected via wired and / or wireless means to enable data transmission between them.

[0072] This invention achieves efficient resource scheduling for big data AI tasks in infectious disease prevention and control through load temperature field modeling, hotspot and coldspot identification, multi-path heat diffusion network construction, task decomposition and resource reservation, dynamic adjustment and execution control. The thermodynamic analogy scheduling method of this invention can accurately identify resource pressure points and optimize paths, effectively balancing computing load and providing an efficient and stable computing support solution for infectious disease prevention and control.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0074] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0075] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A big data AI resource scheduling method for infectious disease prevention and control, characterized in that, include: Obtain real-time load data and infectious disease epidemic situation data for each computing node; Based on the real-time load data, a load temperature field model of the computing node cluster is constructed. By calculating the temperature gradient between adjacent nodes in the load temperature field model, hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold are identified. Obtain the data dependency topology graph of the task to be scheduled, and calculate the transmission energy consumption cost and state synchronization delay penalty of the task migrating between the hot nodes and the cold nodes based on the data dependency topology graph. Based on the transmission energy consumption cost and the state synchronization delay penalty, and combined with the temperature gradient, a multi-path heat dissipation network for task migration is constructed. Based on the multi-path heat diffusion network and the epidemic situation data, a cross-node task decomposition scheme and a spatiotemporal coupling resource reservation strategy are generated. The spatiotemporal coupling resource reservation strategy locks the computing resources of cold node nodes in both the time and space dimensions. By collecting the resource usage pulse sequence of each node during task execution in real time, and performing wavelet transform on the resource usage pulse sequence, the periodic fluctuation pattern and sudden peak characteristics of resource usage are identified. Based on the frequency drift of the periodic fluctuation pattern and the amplitude change of the sudden peak characteristics, the task decomposition granularity in the task decomposition scheme and the node load allocation weight in the multi-path heat diffusion network are dynamically adjusted. Based on the adjusted task decomposition granularity and node load allocation weight, the resource scheduling system is controlled to perform task migration and resource allocation, thereby achieving thermal balance and efficient utilization of computing resources.

2. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The step of constructing a load temperature field model for the computing node cluster based on the real-time load data includes: The CPU utilization, memory utilization, disk I / O throughput, and network I / O throughput of each computing node are obtained to form a node load feature vector; The nodes' load feature vectors are normalized for each dimension, and weight coefficients are assigned according to the degree of influence of different resource types on system performance. Calculate the weighted Euclidean norm of the node load feature vector, and denote it as the node load strength value; Map the node load intensity value to a temperature range and establish a nonlinear mapping function between the load intensity value and the temperature value. The temperature value is assigned to the corresponding computing node, and a discrete temperature distribution is constructed on the spatial topology of the node cluster. Spatial interpolation is performed on the discrete temperature distribution to generate a continuous load temperature field model.

3. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The step of identifying hot spot nodes and cold spot node pairs whose temperature gradients exceed a preset gradient threshold by calculating the temperature gradient between adjacent nodes in the load temperature field model includes: The load temperature value of each computing node in the load temperature field model is smoothed by a time sliding window to eliminate instantaneous load jitter and obtain the smoothed node temperature value. Construct a spatial adjacency graph of the computing node cluster, and determine the set of adjacent nodes for each node based on the spatial adjacency graph; Calculate the difference between the smoothed node temperature value of each node and each node in the set of neighboring nodes, and denote it as the adjacent temperature gradient; Node pairs whose absolute values ​​of adjacent temperature gradients are greater than the preset gradient threshold are recorded as candidate temperature gradient anomaly pairs. For nodes with higher temperature values ​​in the candidate temperature gradient anomaly pair, load persistence verification is performed, and the percentage of time that their temperature values ​​remain at high temperatures within a preset verification duration is calculated and denoted as hotspot stability. Nodes whose hotspot stability is greater than a preset stability threshold are marked as hotspot nodes, and nodes with lower corresponding temperature values ​​are marked as coldspot nodes.

4. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The transmission energy consumption cost and state synchronization latency penalty for the data dependency topology computation task migrating between the hot spot node and the cold spot node include: Analyze the data dependency topology graph to extract the input dataset size and intermediate state data size required for the task; Obtain the network bandwidth and network latency parameters between the hot spot node and the cold spot node; Based on the size of the input dataset, the network bandwidth, and the power consumption per unit of data transmission, calculate the data transmission energy consumption during the task migration process, and record it as the basic transmission energy consumption. Based on the size of the intermediate state data and the memory write rate of the cold node, the reconstruction time cost of the state data at the cold node is calculated and denoted as the state reconstruction latency. The sum of the basic transmission energy consumption and the energy consumption corresponding to the state reconstruction delay is taken as the transmission energy cost; The state synchronization delay penalty is calculated based on the weighted sum of the network latency and the state reconstruction delay, combined with the real-time requirement weight of the task.

5. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The step of constructing a multi-path heat dissipation network for task migration based on the transmission energy consumption cost, the state synchronization delay penalty, and the temperature gradient includes: Using the hot spots as heat sources and the cold spots as cold traps, a heat diffusion topology map is constructed between the nodes. In the heat diffusion topology, a heat diffusion resistance coefficient is calculated for each path from the heat source point to the cold trap point. The heat diffusion resistance coefficient is determined by a normalized weighted average of the transmission energy consumption cost and the state synchronization delay penalty. Calculate the heat transfer efficiency of each path based on the temperature gradient and the thermal diffusion resistance coefficient; Based on the aforementioned heat flux conductivity, a task migration flow quota is allocated to each path; The heat diffusion topology map is integrated with the task migration traffic quota of each path to generate the multi-path heat diffusion network. The multi-path heat diffusion network includes the diversion path and traffic allocation scheme for tasks from hot nodes to multiple cold nodes.

6. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The generation of a cross-node task decomposition scheme and a spatiotemporally coupled resource reservation strategy based on the multi-path heat diffusion network and the epidemic situation data includes: Based on the aforementioned epidemic situation data, predict the growth curve of computing task demand in each region within the future time window; Based on the task migration traffic quota in the multi-path heat diffusion network, determine the target cold spot node set and the task allocation load of each node; Dependency resolution is performed on the computation graph of the task to be migrated, identifying the clusters of subtasks that can be executed in parallel and the critical paths that must be executed serially. Based on the computing power of each node in the target cold point node set and the task allocation load, the subtask cluster is decomposed into multiple subtask units with adaptive computing granularity to generate the task decomposition scheme. Based on the growth curve of the computing task demand and the estimated execution time of the task decomposition scheme, resources are reserved in the target cold node set. The resource reservation simultaneously locks the computing cores, memory capacity and network bandwidth of the nodes, and marks the start time and duration of the reserved resources, thus forming the spatiotemporal coupled resource reservation strategy.

7. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The process of collecting resource usage pulse sequences from each node during task execution in real time, performing wavelet transform on these sequences, and identifying periodic fluctuation patterns and sudden peak characteristics of resource usage includes: The CPU utilization rate, memory utilization rate and network throughput of each node are collected in real time at a preset sampling frequency to form a multi-dimensional resource utilization pulse sequence; Perform wavelet transform on each dimension of the multidimensional resource occupancy pulse sequence to obtain multi-scale wavelet decomposition coefficients; In the multi-scale wavelet decomposition coefficients, low-frequency components are extracted as trend terms of resource occupancy, and mid-frequency components are extracted as candidates for periodic fluctuations. Autocorrelation analysis is performed on the candidate periodic fluctuations to calculate the peak position of the autocorrelation function, and the time interval corresponding to the peak position is recorded as the characteristic period of the periodic fluctuation. The mid-frequency component whose fluctuation amplitude of the characteristic period is greater than the preset period fluctuation threshold is marked as the periodic fluctuation pattern; In the high-frequency components of the multi-scale wavelet decomposition coefficients, amplitude abrupt change points are identified, and points with abrupt change amplitude greater than a preset abrupt change threshold are marked as the burst peak features.

8. The big data AI resource scheduling method for infectious disease prevention and control according to claim 1, characterized in that, The step of dynamically adjusting the task decomposition granularity in the task decomposition scheme and the node load allocation weights in the multipath heat dissipation network based on the frequency drift of the periodic fluctuation pattern and the amplitude changes of the sudden peak characteristics includes: The difference between the current characteristic period and the initial characteristic period of the periodic fluctuation pattern is calculated and denoted as the frequency drift. The frequency and average amplitude of the sudden peak characteristics within a preset statistical window are statistically analyzed and recorded as the sudden intensity index. Based on the direction and amplitude of the frequency drift, the task decomposition granularity adjustment direction is determined. When the frequency drift is positive, the task decomposition granularity is increased; when the frequency drift is negative, the task decomposition granularity is decreased. Based on the aforementioned burst intensity index, calculate the adjustment coefficient for the node load allocation weight; The task decomposition granularity adjustment direction and the adjustment coefficient are applied to the task decomposition scheme to update the calculation granularity of the sub-task unit; Based on the updated calculation granularity and the current load temperature value of each node, the task migration traffic quota of each path in the multipath heat diffusion network is recalculated, and the node load allocation weight is updated.

9. The big data AI resource scheduling method for infectious disease prevention and control according to claim 6, characterized in that, The process involves performing dependency resolution on the computation graph of the task to be migrated, identifying clusters of subtasks that can be executed in parallel and critical paths that must be executed sequentially, including: The computation graph is represented as a directed acyclic graph, where nodes represent computational task units and edges represent data dependencies. Perform topological sorting on the directed acyclic graph to obtain the execution hierarchy sequence of the task units; Within the same execution level, identify a set of task units that do not have data dependencies, and mark this set as the sub-task cluster that can be executed in parallel; From the entry node to the exit node of the directed acyclic graph, calculate the cumulative execution time of all paths, and select the path with the longest cumulative execution time as the critical path. Perform fine-grained analysis on the task units on the critical path to determine whether there is a decomposable data parallel structure within them; The task units with data parallel structure on the critical path are decomposed to generate new sub-task nodes and update the dependencies of the directed acyclic graph.

10. A big data AI resource scheduling system for infectious disease prevention and control, used to implement the big data AI resource scheduling method for infectious disease prevention and control as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire real-time load data and infectious disease epidemic situation data of each computing node; The temperature field modeling module is used to construct a load temperature field model of the computing node cluster based on the real-time load data. The hot spot identification module is used to identify hot spot nodes and cold spot node pairs whose temperature gradient exceeds a preset gradient threshold by calculating the temperature gradient between adjacent nodes in the load temperature field model. The task analysis module is used to obtain the data dependency topology graph of the task to be scheduled, and calculate the transmission energy consumption cost and state synchronization delay penalty of the task migrating between the hot nodes and the cold nodes based on the data dependency topology graph. A heat diffusion network construction module is used to construct a multi-path heat diffusion network for task migration based on the transmission energy consumption cost and the state synchronization delay penalty, combined with the temperature gradient. The task decomposition module is used to generate a cross-node task decomposition scheme and a spatiotemporal coupling resource reservation strategy based on the multi-path heat diffusion network and the epidemic situation data. The spatiotemporal coupling resource reservation strategy locks the computing resources of cold node nodes in both the time and space dimensions. The resource monitoring module is used to collect the resource occupancy pulse sequence of each node in real time during task execution, perform wavelet transform on the resource occupancy pulse sequence, and identify the periodic fluctuation pattern and sudden peak characteristics of resource occupancy. The dynamic adjustment module is used to dynamically adjust the task decomposition granularity in the task decomposition scheme and the node load allocation weight in the multipath heat diffusion network based on the frequency drift of the periodic fluctuation pattern and the amplitude change of the sudden peak characteristics. The execution control module is used to control the resource scheduling system to perform task migration and resource allocation based on the adjusted task decomposition granularity and the node load allocation weight.

Images