Adaptive learning task scheduling system based on multi-objective optimization

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-objective optimization-based adaptive learning task scheduling system, which comprises the following contents: a dependency identification module, an energy consumption analysis module, a bottleneck modeling module, a delay judgment module and an isolation migration module. According to the method, through dependence identification of a task resource calling sequence, acquisition precision of an implicit relationship between tasks is improved, a resource allocation strategy under energy consumption driving is optimized by utilizing an occupancy rate of a node processor, activity of a graphic calculation unit and power supply consumption change, and an I / O interval gravity center shift trend is combined, so that communication congestion prediction capability is enhanced; the abnormal node recognition accuracy is improved by combining the response duration change trend with the node load state, the temperature deviation continuity and the thermal diffusion fluctuation characteristics are used, the thermal imbalance node is sensed in time, task migration is achieved, the node overload rate is effectively reduced, I / O accumulation is relieved, and the continuity and stability of learning task execution are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an adaptive learning task scheduling system based on multi-objective optimization. Background Art

[0002] The field of artificial intelligence technology includes a variety of computing methods and system designs aimed at simulating human intelligent behavior, covering multiple sub-directions such as perception, reasoning, learning and decision-making. The core content includes data-driven model construction, algorithm optimization and task execution mechanism. Relying on mathematical modeling methods such as neural networks, deep learning, and reinforcement learning, it provides solutions to problems such as semantic understanding, image recognition, speech processing, natural language generation, and task planning involved in complex systems. It is applied to scenarios such as autonomous driving, intelligent manufacturing, voice interaction, medical assistance, financial analysis, and cloud computing platform management, and plays an important role in system design such as multi-task collaboration, resource scheduling, and automated optimization. It is composed of a basic algorithm framework, data acquisition and processing mechanism, model training process, system architecture deployment, and dynamic control mechanism.

[0003] Among them, an adaptive learning task scheduling system based on multi-objective optimization refers to a system for dynamically scheduling and allocating resources for multi-task deep learning training processes, covering task division strategies, resource scheduling algorithms, multi-objective optimization mechanisms and adaptive feedback control processes in the deep neural network training process. Specifically, it includes establishing a computational dependency graph for training tasks, quantitatively modeling indicators such as computing time, memory usage, I / O requests, etc. for different tasks, and performing task priority calculations and scheduling order decisions based on a non-dominated sorting genetic algorithm. It introduces a training state evaluation module, dynamically adjusts the task allocation plan according to the model loss change rate, gradient stability and computing node load at each stage, and it iteratively corrects the scheduling results in combination with a constrained local search strategy to form a closed-loop optimization mechanism, control the balance of resource usage and non-interference between tasks, and complete the learning task scheduling configuration of the entire process.

[0004] Traditional learning task scheduling technology mainly relies on statically generated computational dependency graphs and unified resource scheduling algorithms to manage task order and node allocation in training task scheduling. Resource conflicts between tasks are based on logical relationship deduction rather than operational data support, lacking the ability to dynamically analyze the actual access behavior of tasks. Node energy consumption is measured by average resource occupancy, failing to identify the impact of resource combination status on energy consumption changes in different time periods. As a result, the scheduling strategy ignores the resource load concentration problem caused by high-energy-consuming tasks. I / O bottleneck judgment uses I / O request number threshold judgment, lacking early identification of communication congestion trends, resulting in delayed response to task blocking. Response anomaly judgment relies on execution failure or timeout judgment, failing to provide effective early warning for nodes with gradually declining stability. Temperature information is used for fault alarms and is not involved in pre-scheduling decisions. As a result, the system continues to assign new tasks when nodes have poor heat dissipation or heat accumulation trends, increasing the probability of task execution failure and affecting overall stability. This leads to the spread of resource conflicts, node load imbalance, and scheduling lags in a multi-task concurrent environment. Summary of the Invention

[0005] In order to solve the technical problems existing in the prior art, an embodiment of the present invention provides an adaptive learning task scheduling system based on multi-objective optimization. The technical solution is as follows:

[0006] In one aspect, an adaptive learning task scheduling system based on multi-objective optimization is provided, the system comprising:

[0007] The dependency identification module analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, compares the time relationship between each task's calls to the same resource, identifies sequence conflicts, constructs a task dependency graph, and generates dependency information;

[0008] The energy consumption analysis module calls the dependency information, collects the processor usage, graphics computing unit activity and power changes of the node when executing the task, analyzes the energy consumption characteristics corresponding to the computing resource combination, identifies the energy consumption level corresponding to each task, and generates node energy consumption distribution information;

[0009] The bottleneck modeling module calls the node energy consumption distribution information, analyzes the interval change sequence of continuous I / O operations, compares the central trend of the intervals before and after, identifies the interval center of gravity shift behavior, predicts communication congestion based on the change direction, and generates communication congestion trend information;

[0010] The delay determination module uses the communication congestion trend information to compare the processing request response speed of the task in multiple stages, analyzes the changing trend of the average response time in each stage, and combines the changes in the node operation load to screen the nodes and corresponding tasks with abnormal responses and generate the response abnormality processing results.

[0011] As a further solution of the present invention, the dependency information includes resource sequence conflict identification, task dependency path structure, and task call sequence mapping information; the node energy consumption distribution information specifically includes CPU load fluctuation distribution data, GPU activity change pattern, and instantaneous power distribution trend data; the communication congestion trend information includes I / O interval center of gravity offset direction, interval change fluctuation intensity, and continuous request accumulation distribution data; the response exception processing result specifically refers to the average response time change sequence, node load fluctuation characteristic information, and response exception node index.

[0012] As a further solution of the present invention, the dependency identification module includes:

[0013] The resource call extraction submodule analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, screens resource access nodes with continuous call relationships, and establishes a resource call sequence matrix;

[0014] The sequence conflict determination submodule compares the access time sequence relationship of each task to the same resource based on the resource call sequence matrix, detects sequence conflicts, including read-before-write and parallel write, and obtains sequence conflict identification results;

[0015] The dependency path construction submodule selects task pairs with dependency relationships according to the sequence conflict identification result, constructs a task dependency graph structure, and generates dependency relationship information.

[0016] As a further solution of the present invention, the energy consumption analysis module includes:

[0017] The computing resource detection submodule calls the dependency information, collects the processor usage rate, graphics computing unit activity and power change of the node during each task execution, analyzes the combined usage status of the processor and graphics computing unit, and establishes resource combination utilization data;

[0018] The energy consumption level determination submodule analyzes the power consumption changes of the task under various computing resource combination states based on the resource combination utilization data, compares the energy consumption characteristics corresponding to the various states, identifies the energy consumption range of the task during operation, and obtains an energy consumption characteristic determination sequence;

[0019] The distribution information generation submodule identifies the energy consumption level corresponding to each task and generates node energy consumption distribution information by analyzing the energy consumption level distribution of each task in multiple operation stages according to the energy consumption characteristic discrimination sequence.

[0020] As a further solution of the present invention, the specific formula for identifying the energy consumption level corresponding to each task is:

[0021]

[0022] Calculate the energy consumption level of the task;

[0023] Among them, L i Represents the energy consumption level of task i, P′ ij represents the normalized value of node power of task i in the jth stage, U ij represents the activity of the graphics computing unit of task i in the jth stage, M ij represents the processor usage of task i in the jth stage, represents the average processor usage of task i in all stages, K ij represents the load fluctuation impact factor of task i in the jth stage, n represents the total number of running stages of task i, i is the task number index, and j is the task running stage index.

[0024] As a further solution of the present invention, the bottleneck modeling module includes:

[0025] The interval sequence extraction submodule calls the node energy consumption distribution information, collects the start and end time of each I / O operation during the execution of continuous tasks, calculates the time interval between adjacent I / O operations, and establishes an interval sequence;

[0026] The center of gravity shift determination submodule analyzes the density of the preceding and following intervals on the time axis based on the interval sequence, identifies the center of gravity movement direction in the time sequence, and obtains the interval center of gravity shift trend;

[0027] The queuing trend prediction submodule identifies communication queuing behavior based on the interval center of gravity deviation trend and the continuous interval change direction, predicts communication congestion events, and generates communication congestion trend information.

[0028] As a further solution of the present invention, the delay determination module includes:

[0029] The response sequence extraction submodule calls the communication congestion trend information, collects the processing request time and response completion time of the task in multiple stages, calculates the average response time of continuous requests in each stage, and establishes a response time change sequence;

[0030] The change trend identification submodule analyzes the change direction of the average response time in each stage in the time series based on the response time change sequence, identifies the change trend, and obtains the response time change trend by combining the fluctuation amplitude of the response time in the continuous time period;

[0031] The abnormal node screening submodule matches the CPU utilization and I / O load status of the current node according to the response time change trend, and screens the nodes and corresponding tasks with abnormal responses by detecting the synchronization deviation events of utilization changes and response time changes, and maps the tasks to add early warning marks to generate response abnormality processing results.

[0032] As a further solution of the present invention, the specific formula for screening nodes with abnormal responses and corresponding tasks is:

[0033]

[0034] Calculate the response anomaly index for each task;

[0035] Among them, S k is the response abnormality index of task k, C km is the normalized value of CPU utilization of task k in stage m, is the normalized value of the average CPU utilization of task k in all stages, R km is the normalized value of the I / O load state of task k in the mth stage, F km is the hardware adjustment coefficient of task k in the mth stage, D km is the normalized value of the response time of task k in the mth stage, E km is the normalized value of the node load intensity of task k in the mth stage, G km is the load intensity influencing factor of task k in the mth stage, M is the total number of running stages of task k, k is the task number, and m is the number of the mth stage during the execution of task k.

[0036] As a further embodiment of the present invention, the system further comprises:

[0037] The isolation migration module calls the response exception processing result, collects temperature data of each node during continuous operation, and compares it with the preset operating reference temperature range, identifies and analyzes the persistence of temperature offsets, and detects abnormal heat diffusion events based on the fluctuation characteristics of the temperature data. The corresponding tasks are split and migrated to multiple nodes, and a thermal migration adjustment record is generated;

[0038] The thermal migration adjustment record specifically includes a temperature continuous offset sequence, a thermal diffusion anomaly identification label, and a task split migration path mapping result.

[0039] As a further solution of the present invention, the isolation migration module includes:

[0040] The temperature offset detection submodule calls the response exception processing result, collects the temperature sequence of each node during continuous operation, analyzes the offset state of each temperature value relative to the node operation reference temperature range, analyzes the persistence of the temperature offset, and establishes a temperature offset change sequence;

[0041] The heat diffusion anomaly determination submodule selects the node state of heat diffusion imbalance based on the temperature offset change sequence and the temperature fluctuation characteristics in the continuous time window, detects heat diffusion anomaly events, and obtains the heat diffusion anomaly trend indicator;

[0042] The task migration mapping submodule extracts the nodes and tasks corresponding to the abnormal event based on the heat diffusion abnormal trend indicator, splits the corresponding tasks, and migrates the split tasks to multiple nodes according to the resource occupancy status of each node to generate heat-sensitive migration adjustment records.

[0043] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0044] By identifying the dependencies on the order of task resource calls, the accuracy of capturing implicit relationships between tasks is improved. By utilizing the node processor occupancy rate, graphics computing unit activity and power consumption changes, the energy-driven resource allocation strategy is optimized. Combined with the I / O interval center of gravity shift trend, the communication congestion prediction capability is enhanced. The response time change trend is combined with the node load status to improve the accuracy of abnormal node identification. The temperature offset persistence and heat diffusion fluctuation characteristics are used to achieve timely perception of thermally unbalanced nodes and task migration, effectively reducing the node overload rate, alleviating I / O accumulation, and improving the continuity and stability of learning task execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 is a system flow chart of the present invention;

[0047] Figure 2 Schematic diagram of the system framework of the present invention;

[0048] Figure 3 This is a flow chart of the dependency identification module of the present invention;

[0049] Figure 4 This is a flow chart of the energy consumption analysis module of the present invention;

[0050] Figure 5This is a flowchart of the bottleneck modeling module of the present invention;

[0051] Figure 6 This is a flow chart of the delay determination module of the present invention;

[0052] Figure 7 This is a flow chart of the isolation migration module of the present invention. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0054] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0055] In the embodiments of the present invention, the terms "image" and "picture" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same. The terms "of," "corresponding," and "corresponding" may be used interchangeably. It should be noted that, when the distinction between them is not emphasized, their intended meanings are the same.

[0056] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0057] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0058] The embodiment of the present invention provides an adaptive learning task scheduling system based on multi-objective optimization, see Figures 1 to 2 The present invention provides a technical solution, an adaptive learning task scheduling system based on multi-objective optimization, comprising:

[0059] The dependency identification module analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, compares the time relationship between each task's calls to the same resource, identifies sequence conflicts, constructs a task dependency graph, and generates dependency information;

[0060] The energy consumption analysis module calls dependency information to collect the processor usage, graphics computing unit activity, and power changes of the node when executing tasks. It analyzes the energy consumption characteristics corresponding to the computing resource combination, identifies the energy consumption level corresponding to each task, and generates node energy consumption distribution information.

[0061] The bottleneck modeling module uses node energy consumption distribution information to analyze the interval change sequence of continuous I / O operations, compares the central trend of the intervals before and after, identifies the interval center of gravity offset behavior, and predicts communication congestion based on the change direction to generate communication congestion trend information;

[0062] The delay determination module uses communication congestion trend information to compare the response speed of task processing requests in multiple stages, analyze the changing trend of the average response time in each stage, and screen the nodes and corresponding tasks with abnormal responses based on the changes in node operating load to generate response abnormality processing results.

[0063] The isolation migration module calls the response exception processing results, collects the temperature data of each node during continuous operation, and compares it with the preset operating reference temperature range, identifies and analyzes the persistence of temperature offsets, and detects abnormal heat diffusion events based on the fluctuation characteristics of temperature data. The corresponding tasks are split and migrated to multiple nodes to generate thermal migration adjustment records.

[0064] Dependency information includes resource sequence conflict identification, task dependency path structure, and task call sequence mapping information. Node energy consumption distribution information specifically includes CPU load fluctuation distribution data, GPU activity change pattern, and instantaneous power distribution trend data. Communication congestion trend information includes I / O interval center of gravity offset direction, interval change fluctuation intensity, and continuous request accumulation distribution data. Response exception processing results specifically refer to the average response time change sequence, node load fluctuation feature information, and response exception node index. Thermal-sensitive migration adjustment records specifically include temperature continuous offset sequence, thermal diffusion anomaly identification label, and task split migration path mapping results.

[0065] See also Figure 2 and Figure 3 , the dependency identification module includes:

[0066] The resource call extraction submodule analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, screens resource access nodes with continuous call relationships, and establishes a resource call sequence matrix;

[0067] The resource call extraction submodule analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages. It is necessary to obtain the memory address A1 corresponding to the loading of model weight data in the initialization stage of the training task, the cache address A2 when performing matrix operations in the calculation stage, and the write address A3 accessed when updating parameters in the return stage. A1, A2, and A3 represent the memory locations called by the task in the three stages respectively. The corresponding access timestamps T1, T2, and T3 are recorded in the order of the system clock for each address call, where T1, T2, and T3 represent the time of accessing the corresponding memory address. Mark, and extract the I / O channel numbers P1, P2, and P3 used by the task in these three stages, where P1, P2, and P3 represent the transmission path identifiers assigned to each stage, mark the calling channels and time order, establish a resource access record table in the task number dimension, and form a basic data set of the task and resource calling sequence. For all access events in the set, determine whether there is a continuity condition within the access cycle. Continuity is defined as the difference between the latter access timestamp and the previous access timestamp is less than the set maximum interval threshold ΔT, where ΔT represents the maximum allowable time difference between consecutive accesses. If T i+1 -T i ≤ΔT, where T i+1 is the last access timestamp, T i If it is the timestamp of the previous access, the corresponding resource number is classified as a continuous call node, and all access event pairs that meet the continuity conditions are arranged in chronological order to construct the access node sequence V i , where V i Represents the access combination of the i-th pair of consecutively called resources, and marks the resource identifiers and stage attributes involved in the sequence, and aggregates the resource address pairs accessed continuously in the same task number into a stage access chain L by stage. k , where L k Represents the kth continuous access stage in the task, and uses the task number as the index to construct a three-dimensional matrix S(t,r,p), where t represents the task number, r represents the resource number, and p represents the stage the task is in. Each element S(t,r,p) in the matrix records the continuous access relationship of task t to resource r in stage p. If there are adjacent access events with the same resource number but different stage labels, and the time is continuous, the matrix value is marked as 1, otherwise it is 0. Finally, the matrix is ​​aggregated according to the task number dimension to establish the call sequence relationship matrix between tasks and resources in different stages, and generate the resource call sequence matrix.

[0068] The sequence conflict determination submodule compares the access time sequence of each task to the same resource based on the resource call sequence matrix, detects sequence conflicts, including read-before-write and parallel write, and obtains sequence conflict identification results;

[0069] Based on the resource call order matrix, the sequential conflict determination sub-module needs to compare the resource access times of all elements marked as consecutive calls in the matrix. First, extract the access timestamps T1(t1, r, p1) and T2(t2, r, p2) of the same resource r under task numbers t1 and t2 in their respective stages p, where t1 and t2 are two different task numbers, r is the resource number, p1 and p2 are stage identifiers, and T1 and T2 are the access timestamps of the two tasks for resource r. Perform the T1 < T2 and access type matching judgment. When t1 performs a write operation on r and t2 performs a read operation on r, it satisfies the write-before-read conflict condition. If both t1 and t2 are write operations and T1 ≈ T, the parallel write conflict condition is satisfied. The approximate relationship sets an error range of ε ≤ 5 ms, where ε represents the minimum time interval tolerance set by the scheduler. If T2 - T1 ≤ ε and the operation types are the same, it is marked as a parallel conflict. Record the task number pair (t1, t2), the conflict resource number r, and the conflict type for all task combinations that satisfy the conflict conditions, and construct a conflict record table C i , where C<0000​​​​​​​​​​Represents the interlocking relationship that needs to be processed synchronously between tasks, maps the task number to a graph node, and constructs an initial graph structure G(V,E) for all task nodes, where V is the task set and E is the dependent edge set. Then, the graph structure is connected. For task pairs that share resources but are not connected by conflict events, determine whether there is an indirect dependency path. If a graph search finds that a path from t1 to t2 exists but is not unique, perform shortest path reconstruction, set the edge weight to the access order difference ΔT = T(t2) - T(t1), where ΔT represents the difference in resource access time between the two tasks, and use the path with the minimum weight as the valid dependent path and add it to the edge set. Perform topological sorting on the final graph structure G to verify whether there is a ring structure. If there is a directed ring between the task pairs, mark the ring task set as a resource interlocking set and record it as a strong dependency constraint unit in the task scheduling process. Record all edge weights and dependency direction information as the structure vector D i , where D i Represents the i-th dependency path data unit. The elements in the vector contain the source task number, target task number, resource number and edge type label. All vector data are summarized to generate dependency information.

[0072] See also Figure 2 and Figure 4 , the energy consumption analysis module includes:

[0073] The computing resource detection submodule calls dependency information, collects the node's processor usage, graphics computing unit activity, and power changes during each task execution, analyzes the combined usage status of the processor and graphics computing unit, and establishes resource combination utilization data;

[0074] The computing resource detection submodule calls dependency information to collect the node's processor usage, graphics computing unit activity, and power changes during each task run. The processor occupancy percentage U is recorded from the start time to the end time of the task corresponding to the task number t. cpu (t), Graphics computing unit usage U gpu (t) and the average power reading P(t), where U cpu (t) represents the real-time CPU utilization of task t during execution, U gpu (t) represents the proportion of GPU parallel core activity, P(t) represents the power consumption statistics per unit time during the task execution phase, and the three data items in each sampling period are obtained to form a time series vector group The superscript i represents the i-th sampling period, and is arranged as a vector in each time period and stored in a three-dimensional matrix R(t,i,j), where t represents the task number, i is the time series index, and j is the resource type index. j = 1 represents CPU, j = 2 represents GPU, and j = 3 represents power consumption value. Each group of i rows in the matrix is ​​combined with a label to determine whether the task is CPU-dominated, GPU-dominated, or mixed execution state in the time period. and The relative proportion of When it is defined as the CPU-dominated segment, if it is less than 0.3, it is the GPU-dominated segment, and the rest of the interval is a mixed segment. The segment type is assigned to the combined state sequence C(t,i), and finally combined with the power consumption value P i Associated with the combination state sequence, the resource usage classification statistics under the task number dimension are established to generate resource combination utilization data.

[0075] The energy consumption level determination submodule analyzes the power consumption changes of tasks under various computing resource combination states based on resource combination utilization data, compares the energy consumption characteristics corresponding to various states, identifies the energy consumption range of tasks during operation, and obtains the energy consumption characteristic judgment sequence;

[0076] The energy consumption level determination submodule analyzes the power consumption changes of tasks under various computing resource combination states based on resource combination utilization data, and compares the combination type sequence C(t,i) and the corresponding power consumption value P under each task number t. i Combine and calculate the average power consumption data of all CPU-dominated sections GPU-dominated segment calculation mean The average value of the mixed section is in Among them S cpu ,S gpu ,S mix are the sampling sequence index sets corresponding to the CPU-dominated segment, GPU-dominated segment, and mixed segment, respectively. n, m, and k are the number of samples in each segment, and the sample capacity under each resource usage state. After calculating the average energy consumption level under the three segments, further analyze its fluctuation trend within the task cycle, perform linear fitting on the continuous power consumption sequence of each segment, and extract the fitting slope k. i , judge whether the power consumption trend is stable, increasing or decreasing, label the trend under each type of combination, establish a state mapping table for the power consumption mean and trend combination, form a description of the power consumption level of the task under each combination resource mode, and finally construct a combined energy consumption expression sequence under the task number dimension to obtain the energy consumption characteristic discrimination sequence.

[0077] The distribution information generation submodule discriminates the sequence according to the energy consumption characteristics, analyzes the energy consumption level distribution of each task in multiple operation stages, identifies the energy consumption level corresponding to each task, and generates node energy consumption distribution information;

[0078] The specific formula for identifying the energy consumption level corresponding to each task is:

[0079]

[0080] Calculate the energy consumption level of the task;

[0081] Among them, L i Represents the energy consumption level of task i, P′ ij represents the normalized value of node power of task i in the jth stage, U ij represents the activity of the graphics computing unit of task i in the jth stage, M ij represents the processor usage of task i in the jth stage, represents the average processor usage of task i in all stages, K ij represents the load fluctuation impact factor of task i in the jth stage, n represents the total number of running stages of task i, i is the task number index, and j is the task running stage index.

[0082] formula:

[0083]

[0084] Detailed explanation of the formula and the process of formula calculation and derivation:

[0085] The formula is used to calculate the energy consumption level of task i in n running stages. The results are used to determine the energy consumption status of the node and support task scheduling optimization.

[0086] Parameter meaning and setting value:

[0087] P′ ij The normalized value of the node power of task i in the jth stage is calculated based on the node's current stage power value P and the maximum power Pmax during the node power measurement period. In the current stage, P is 175 watts and Pmax is 250 watts. The calculated value of Pij' is 0.7, which reflects the ratio of the node's current stage power consumption to the maximum power.

[0088] U ij The GPU activity of task i in the jth stage is set to 0.82, which reflects the GPU activity ratio.

[0089] M ij is the processor usage of task i in the jth stage, and the current collection value is set to 0.78, reflecting the CPU load level;

[0090] is the average processor usage of task i in all stages. Assuming the total number of current stages is 3, the CPU usage of the three stages are 0.78, 0.65, and 0.71 respectively. The calculation results are

[0091] K ij is the load fluctuation impact factor of task i in the jth stage. It is adjusted according to the fluctuation sensitivity. If the CPU usage fluctuates by more than 0.05 compared with the average value, K is set to 1.2; otherwise, it is set to 1.0. In this stage, |0.78-0.713|=0.067 is greater than 0.05, so K is 1.2.

[0092] n is the total number of running stages of task i, and the current number of stages is set to 3;

[0093] Substitute the parameters into the formula for calculation:

[0094] Calculate the sub-items for each stage:

[0095] Phase 1:

[0096]

[0097] Phase 2:

[0098] Set the node power normalization value to 0.6, GPU activity to 0.75, and CPU usage to 0.65. The fluctuation is |0.65-0.713|=0.063, which is greater than 0.05, and K is 1.2

[0099] (0.6 0.75) + (0.063 1.2) = 0.45 + 0.0756 = 0.5256;

[0100] Phase 3:

[0101] Set the node power normalization value to 0.68, GPU activity to 0.8, CPU usage to 0.71, fluctuation |0.71-0.713|=0.003 less than 0.05, and K to 1.0

[0102] (0.68·0.8)+(0.003·1.0)=0.544+0.003=0.547;

[0103] Substitute into the formula for calculation:

[0104] 0.6544+0.5256+0.547=1.727;

[0105]

[0106] The result 0.5757 indicates that the energy consumption level of task i in the current operation cycle is 0.5757. This value reflects the comprehensive energy consumption level of the task under the CPU load fluctuation, GPU activity and node power usage status. This value will be used as an energy consumption level parameter for node energy consumption distribution mapping, providing a basis for subsequent task scheduling.

[0107] See also Figure 2 and Figure 5 , the bottleneck modeling module includes:

[0108] The interval sequence extraction submodule calls the node energy consumption distribution information, collects the start and end time of each I / O operation during the execution of continuous tasks, calculates the time interval between adjacent I / O operations, and establishes the interval sequence;

[0109] The interval sequence extraction submodule calls the node energy consumption distribution information, collects the start and end time of each I / O operation during the execution of the continuous task, and records the start time S of all I / O requests of task number t in the execution phase. i and end time E i , where S i Indicates the starting timestamp of the i-th I / O operation, E i Indicates the completion timestamp of the corresponding operation, and calculates the interval between two consecutive operations G i =S i+1 -E i , where G i Represents the time interval between the i-th and i+1-th I / O operations, for all G i Arrange in time order to form an interval sequence G={G1,G2,…,G n}, save the sequence with the task number t as the index mapping, mark the execution period of each interval, divide the task execution period into equal time windows, accumulate the number of intervals in each window, and form the interval density distribution vector D = {D1, D2, ..., D k}, where D k It represents the number of I / O intervals in the kth time window. The window length is set to 20 seconds. This value is set according to the standard statistical period of the operating system. Finally, the task number and its interval sequence sequence are indexed and mapped, recorded as task-level time interval dynamic data, and the interval sequence sequence is generated.

[0110] The center of gravity shift determination submodule analyzes the density of the previous and next intervals on the time axis based on the interval sequence, identifies the center of gravity movement direction in the time series, and obtains the interval center of gravity shift trend;

[0111] The center of gravity offset determination submodule analyzes the density of the preceding and following intervals on the time axis based on the interval sequence, and converts the interval density distribution vector D = {D1, D2, ..., D k}Use the center of gravity formula to calculate the center of gravity position C, where Where i represents the time window index, D i is the number of I / O intervals in the window, and the calculated C is the position label of the density center of gravity on the time axis. The current task execution cycle is divided into five equal-length segments. If the segment position of C is less than 3, it is marked as the front-biased center of gravity, equal to 3 as the center of gravity, and greater than 3 as the back-biased center of gravity. In this way, the interval center of gravity offset trend label T is identified, where T is the center of gravity offset classification result of the task in the current execution cycle. Then, it is determined whether C has changed compared with the previous task cycle, and the direction of center of gravity change is extracted. If the current round C is greater than the center of gravity value C of the previous cycle, prev , it is defined as a backward trend. If it is less than, it is a forward trend. Finally, the offset direction and the task time period structure identifier are combined to record the dynamic position change trend of the center of gravity on the time axis to obtain the interval center of gravity offset trend.

[0112] The queuing trend prediction submodule identifies communication queuing behavior based on the interval center of gravity offset trend and the direction of change of continuous intervals, predicts communication congestion events, and generates communication congestion trend information;

[0113] The queue trend prediction submodule makes a judgment based on the interval center of gravity deviation trend and the continuous interval change direction. First, the interval sequence G={G1,G2,…,G n Perform first-order difference processing to calculate the change ΔG between adjacent intervals i =G i+1 -G i , where ΔG i Indicates the increasing and decreasing trend of I / O interval between consecutive requests. The number of positive and negative values ​​in the differential sequence is extracted. If there are three or more consecutive ΔG i >0, the period is marked as a continuously lengthened interval, corresponding to the judgment of queue growth behavior, and the corresponding value is a delay accumulation signal. In the obtained center of gravity shift trend label T, if the value is a backward trend and the continuous value-added ratio in the differential sequence exceeds 70%, the task is combined and marked as a communication congestion risk task, and the current congestion level label L is calculated. q , where L q =3 means high risk, L q =2 means medium risk, L q =1 indicates low risk, and the judgment benchmark is set as the number of consecutive positive growth segments under high risk ≥ 3 and the maximum interval G iIf the value exceeds 1.5 times the maximum value of the first 50% segment, the task number, risk level, maximum interval value, and change direction that meet the conditions will be recorded to generate communication congestion trend information.

[0114] See also Figure 2 and Figure 6 , the delay determination module includes:

[0115] The response sequence extraction submodule calls for communication congestion trend information, collects task processing request time and response completion time in multiple stages, calculates the average response time of consecutive requests in each stage, and establishes a response time change sequence;

[0116] The response sequence extraction submodule calls the communication congestion trend information, collects the processing request time and response completion time of the task in multiple stages, and extracts the request initiation time of all processing requests in stage number p in task number t Response end time in Indicates the timestamp of the initiation of the i-th request, Indicates the timestamp of the corresponding request response completion and calculates the response duration The average of all response times in each stage is calculated to form a stage-level response time mean sequence. in n is the number of requests in a stage, and the average response time of all stages is recorded according to the task number to form a task response time series. Where m is the total number of stages in which the task is divided. The stage index and time window corresponding to each item are marked. The sequence is archived uniformly according to the task number index to generate a response time change sequence.

[0117] The change trend identification submodule analyzes the change direction of the average response time in each stage in the time series based on the response time change sequence, identifies the change trend, and combines the fluctuation amplitude of the response time in the continuous time period to obtain the response time change trend;

[0118] The change trend identification submodule is based on the response time change sequence, and analyzes the change direction of the average response time in each stage in the time series. Perform first-order difference calculations Where j represents the stage index. If three consecutive difference values ​​are positive, it is defined as an increasing trend. If they are consecutively negative, it is defined as a decreasing trend. The others are fluctuating trends. At the same time, for all the differences in each stage, Calculating standard deviation where σ (p)Represents the response fluctuation amplitude of stage p. If the standard deviation in three consecutive stages is greater than 40% of the mean, it is marked as a fluctuation interval. The trend direction and the fluctuation label are combined and recorded to generate the response sequence trend vector V of the corresponding task. r (t), record the trend label, change direction and fluctuation judgment of each stage, and finally summarize it to form the response duration change trend.

[0119] The abnormal node screening submodule matches the CPU utilization and I / O load status of the current node based on the response time change trend. By detecting the synchronization deviation events between utilization changes and response time changes, it screens the nodes and corresponding tasks with abnormal responses, maps the tasks, adds warning tags, and generates response abnormality processing results.

[0120] The specific formula for filtering nodes with abnormal responses and corresponding tasks is:

[0121]

[0122] Calculate the response anomaly index for each task;

[0123] Among them, S k is the response abnormality index of task k, C km is the normalized value of CPU utilization of task k in stage m, is the normalized value of the average CPU utilization of task k in all stages, R km is the normalized value of the I / O load state of task k in the mth stage, F km is the hardware adjustment coefficient of task k in the mth stage, D km is the normalized value of the response time of task k in the mth stage, E km is the normalized value of the node load intensity of task k in the mth stage, G km is the load intensity influencing factor of task k in the mth stage, M is the total number of running stages of task k, k is the task number, and m is the number of the mth stage during the execution of task k.

[0124] formula:

[0125]

[0126] Detailed explanation of the formula and the process of formula calculation and derivation:

[0127] This formula is used to calculate the response anomaly index of task k, reflecting whether there are abnormalities in the response of the task at different stages;

[0128] Parameter meaning and setting value:

[0129] S k: The response anomaly index of task k, indicating whether task k has an abnormal response. A larger value indicates a higher degree of anomaly.

[0130] C km : The normalized value (unitless) of the CPU utilization of task k in the mth stage, calculated by monitoring the ratio of CPU usage time to total time, with a value range of 0 to 1;

[0131] The normalized value (unitless) of the average CPU utilization of task k across all stages, calculated by averaging the CPU utilization across all stages;

[0132] R km : The normalized value (unitless) of the I / O load status of task k in the mth stage, calculated as the ratio of the number of I / O requests to the total requests, ranging from 0 to 1;

[0133] F km : The hardware adjustment coefficient of task k in the mth stage (unitless), adjusted according to the hardware performance requirements, the value is generally between 1 and 2;

[0134] D km : The normalized value (unitless) of the response time of task k in the mth stage, calculated by monitoring the ratio of the time to complete the task to the total time, with a value range of 0 to 1;

[0135] E km : The normalized value of the node load intensity of task k in the mth stage (unitless), calculated by monitoring the ratio of the node power consumption to the total power consumption, with a value range of 0 to 1;

[0136] G km : The load intensity impact factor of task k in the mth stage (unitless), which is dynamically adjusted according to the workload intensity of the node and its impact on task response;

[0137] M: The total number of running stages of task k, which indicates the number of stages the task goes through from start to finish;

[0138] Assume that task k has gone through 3 stages, that is, M = 3, C k1 =0.7, C k2 =0.8, C k3 =0.75 (normalized value of CPU usage of tasks at each stage), (normalized average CPU usage of task k), R k1 =0.9, R k2 =0.8, R k3 =0.85 (normalized value of the I / O load state of the task at each stage), F k1 =1.1, F k2=1.2, F k3 =1.3 (hardware adjustment coefficient), D k1 =0.6, D k2 =0.7, D k3 =0.65 (normalized response time), E k1 =0.5, E k2 =0.6, E k3 =0.55 (normalized value of node load intensity), G k1 =1.1, G k2 =1.15, G k3 =1.2 (load intensity influence factor);

[0139] Substitute the parameters into the formula for calculation:

[0140] Calculate the bias and weighted influence of each stage:

[0141]

[0142] Calculate the weighted impact of response time and load intensity:

[0143]

[0144]

[0145] Calculate the response anomaly index:

[0146]

[0147] The calculated S k = 0.0614, indicating that task k has relatively low response anomalies across all phases. This value indicates that the task's CPU utilization, I / O load, and hardware adjustment coefficients vary slightly across phases, and the combined impact of response time and node load is relatively balanced, with no significant response anomalies.

[0148] See also Figure 2 and Figure 7 , the isolated migration module includes:

[0149] The temperature offset detection submodule calls and responds to the abnormal processing results, collects the temperature sequence of each node during continuous operation, analyzes the offset status of each temperature value relative to the node's operating reference temperature range, analyzes the persistence of the temperature offset, and establishes a temperature offset change sequence;

[0150] The temperature offset detection submodule calls the response exception processing result, collects the temperature sequence of each node during continuous operation, and records the core temperature of node number n every minute during the operation cycle. Where i represents the sampling time series index, and a temperature sampling sequence is established. At the same time, set the reference temperature range of the node [T min ,T max ], where T min With T max The recommended minimum and maximum core temperatures for the hardware operating environment are 40°C and 85°C respectively. Execute offset judgment, if or Then mark the item as an offset point, further count the consecutive offset points, and determine whether 3 or more consecutive time points are in the offset state. If so, mark it as a continuous offset segment, and record the start and end indexes s, e of the segment. Calculate the maximum-minimum difference D of the temperature fluctuation amplitude of each continuous offset segment. (j) =max(T (j) )-min(T (j) ), where j represents the jth duration, if D (j) If the temperature is less than 3℃, it is considered as a stable offset segment. The node number, offset segment index, and fluctuation amplitude are archived, and a node temperature offset trend list is constructed to generate a temperature offset change sequence.

[0151] The heat diffusion anomaly determination submodule screens the node states with heat diffusion imbalance based on the temperature offset change sequence and the temperature fluctuation characteristics within the continuous time window, detects heat diffusion anomaly events, and obtains the heat diffusion anomaly trend indicator;

[0152] The heat diffusion anomaly judgment submodule divides the node operation cycle into multiple time windows W according to the temperature offset change sequence and the temperature fluctuation characteristics in the continuous time window. k , each window length is set to 5 minutes, and the temperature sequence in each window Calculate its standard deviation Where m is the number of samples in the window, is the mean temperature in the window, if σ k If the temperature is less than 1.5℃ and there are at least three consecutive offset points in the corresponding window, the window is marked as a slow heat diffusion section. A fluctuation detection threshold sequence is set for all windows. The window combination that meets this condition is used to establish an index set H = {W1, W4, ...}. The nodes to which the windows belong in the set are clustered and identified. The node is determined to be a heat diffusion imbalance node. The heat diffusion trend level index is then calculated based on its consecutive occurrence times and window proportion. If E n A value ≥0.4 is defined as a significant heat diffusion anomaly. The node number, number of anomaly windows, and diffusion trend level are summarized and archived to generate a heat diffusion anomaly trend index.

[0153] The task migration mapping submodule extracts the nodes and tasks corresponding to abnormal events based on the heat diffusion anomaly trend indicator, splits the corresponding tasks, and migrates the split tasks to multiple nodes based on the resource usage status of each node, generating heat-sensitive migration adjustment records;

[0154] The task migration mapping submodule extracts the node set N corresponding to the abnormal event based on the heat diffusion abnormal trend indicator. a With the task set T a , extract the task and node binding map M(t,n) according to the task allocation record, and for each task t∈T a Perform granular splitting and divide the task into subtasks t1, t2, ..., t according to the computational load partition in the task execution record. m At the same time, the system current resource monitoring module is called to obtain the CPU and GPU resource occupancy rates of all non-abnormal nodes, and the resource utilization sequence R is constructed. j ={C j ,G j}, where C j is the CPU usage of node j, G j For its GPU activity rate, the nodes whose resource usage is lower than the set scheduling benchmark are screened into the candidate node set N c , set the scheduling benchmark to less than 60% for CPU and less than 50% for GPU, and match each subtask to N in order according to the resource consumption level. c The node with the best idleness updates the mapping relationship M(t i ,n j ), complete the task redistribution and record the new allocation table, record the source node, target node, migration subtask number and scheduling time window for each migration operation, and finally generate a hot-sensitive migration adjustment record.

[0155] The above embodiments can be implemented in whole or in part through software, hardware (such as circuits), firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the processes or functions described in accordance with the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired method (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, or magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0156] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0157] In this disclosure, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, "at least one of a, b, or c" can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0158] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0159] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0160] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0161] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of the device or unit, which can be electrical, mechanical or other forms.

[0162] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0163] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0164] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0165] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An adaptive learning task scheduling system based on multi-objective optimization, characterized in that: The system comprises: The dependency identification module analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, compares the time relationship between each task's calls to the same resource, identifies sequence conflicts, constructs a task dependency graph, and generates dependency information; The energy consumption analysis module calls the dependency information, collects the processor usage, graphics computing unit activity and power changes of the node when executing the task, analyzes the energy consumption characteristics corresponding to the computing resource combination, identifies the energy consumption level corresponding to each task, and generates node energy consumption distribution information; The bottleneck modeling module calls the node energy consumption distribution information, analyzes the interval change sequence of continuous I / O operations, compares the central trend of the intervals before and after, identifies the interval center of gravity shift behavior, predicts communication congestion based on the change direction, and generates communication congestion trend information; The delay determination module uses the communication congestion trend information to compare the processing request response speed of the task in multiple stages, analyzes the changing trend of the average response time in each stage, and combines the changes in the node operation load to screen the nodes and corresponding tasks with abnormal responses and generate the response abnormality processing results.

2. The adaptive learning task scheduling system based on multi-objective optimization according to claim 1 is characterized in that: The dependency information includes resource sequence conflict identifier, task dependency path structure, and task call sequence mapping information; the node energy consumption distribution information specifically includes CPU load fluctuation distribution data, GPU activity change pattern, and instantaneous power distribution trend data; the communication congestion trend information includes I / O interval center of gravity offset direction, interval change fluctuation intensity, and continuous request accumulation distribution data; the response exception processing result specifically refers to the average response time change sequence, node load fluctuation characteristic information, and response exception node index.

3. The adaptive learning task scheduling system based on multi-objective optimization according to claim 1 is characterized in that: The dependency identification module includes: The resource call extraction submodule analyzes the memory address call sequence and I / O channel access sequence corresponding to each training task in multiple stages, screens resource access nodes with continuous call relationships, and establishes a resource call sequence matrix; The sequence conflict determination submodule compares the access time sequence relationship of each task to the same resource based on the resource call sequence matrix, detects sequence conflicts, including read-before-write and parallel write, and obtains sequence conflict identification results; The dependency path construction submodule selects task pairs with dependency relationships according to the sequence conflict identification result, constructs a task dependency graph structure, and generates dependency relationship information.

4. The adaptive learning task scheduling system based on multi-objective optimization according to claim 3 is characterized in that: The energy consumption analysis module includes: The computing resource detection submodule calls the dependency information, collects the processor usage rate, graphics computing unit activity and power changes of the node during each task execution, analyzes the combined usage status of the processor and graphics computing unit, and establishes resource combination utilization data; The energy consumption level determination submodule analyzes the power consumption changes of the task under various computing resource combination states based on the resource combination utilization data, compares the energy consumption characteristics corresponding to the various states, identifies the energy consumption range of the task during operation, and obtains an energy consumption characteristic determination sequence; The distribution information generation submodule identifies the energy consumption level of each task in multiple operation stages according to the energy consumption characteristic discrimination sequence, identifies the energy consumption level corresponding to each task, and generates node energy consumption distribution information.

5. The adaptive learning task scheduling system based on multi-objective optimization according to claim 4 is characterized in that: The specific formula for identifying the energy consumption level corresponding to each task is: Calculate the energy consumption level of the task; Among them, L i Represents the energy consumption level of task i, P′ ij represents the normalized value of node power of task i in the jth stage, U ij represents the activity of the graphics computing unit of task i in the jth stage, M ij represents the processor usage of task i in the jth stage, represents the average processor usage of task i in all stages, K ij represents the load fluctuation impact factor of task i in the jth stage, n represents the total number of running stages of task i, i is the task number index, and j is the task running stage index.

6. The adaptive learning task scheduling system based on multi-objective optimization according to claim 4 is characterized in that: The bottleneck modeling module includes: The interval sequence extraction submodule calls the node energy consumption distribution information, collects the start and end time of each I / O operation during the execution of continuous tasks, calculates the time interval between adjacent I / O operations, and establishes an interval sequence; The center of gravity shift determination submodule analyzes the density of the preceding and following intervals on the time axis based on the interval sequence, identifies the center of gravity movement direction in the time sequence, and obtains the interval center of gravity shift trend; The queuing trend prediction submodule identifies communication queuing behavior based on the interval center of gravity deviation trend and the continuous interval change direction, predicts communication congestion events, and generates communication congestion trend information.

7. The adaptive learning task scheduling system based on multi-objective optimization according to claim 6 is characterized in that: The delay determination module includes: The response sequence extraction submodule calls the communication congestion trend information, collects the processing request time and response completion time of the task in multiple stages, calculates the average response time of continuous requests in each stage, and establishes a response time change sequence; The change trend identification submodule analyzes the change direction of the average response time in each stage in the time series based on the response time change sequence, identifies the change trend, and obtains the response time change trend by combining the fluctuation amplitude of the response time in the continuous time period; The abnormal node screening submodule matches the CPU utilization and I / O load status of the current node according to the response time change trend, and screens the nodes and corresponding tasks with abnormal responses by detecting the synchronization deviation events of utilization changes and response time changes, and maps the tasks to add early warning marks to generate response abnormality processing results.

8. The adaptive learning task scheduling system based on multi-objective optimization according to claim 7 is characterized in that: The specific formula for filtering nodes with abnormal responses and corresponding tasks is: Calculate the response anomaly index for each task; Among them, S k is the response abnormality index of task k, C km is the normalized value of CPU utilization of task k in stage m, is the normalized value of the average CPU utilization of task k in all stages, R km is the normalized value of the I / O load state of task k in the mth stage, F km is the hardware adjustment coefficient of task k in the mth stage, D km is the normalized value of the response time of task k in the mth stage, E km is the normalized value of the node load intensity of task k in the mth stage, G km is the load intensity influencing factor of task k in the mth stage, M is the total number of running stages of task k, k is the task number, and m is the number of the mth stage during the execution of task k.

9. The adaptive learning task scheduling system based on multi-objective optimization according to claim 1 is characterized in that: The system further comprises: The isolation migration module calls the response exception processing result, collects temperature data of each node during continuous operation, and compares it with the preset operating reference temperature range, identifies and analyzes the persistence of temperature offsets, and detects abnormal heat diffusion events based on the fluctuation characteristics of the temperature data. The corresponding tasks are split and migrated to multiple nodes, and a thermal migration adjustment record is generated; The thermal migration adjustment record specifically includes a temperature continuous offset sequence, a thermal diffusion anomaly identification label, and a task split migration path mapping result.

10. The adaptive learning task scheduling system based on multi-objective optimization according to claim 9, characterized in that: The isolation migration module includes: The temperature offset detection submodule calls the response exception processing result, collects the temperature sequence of each node during continuous operation, analyzes the offset state of each temperature value relative to the node operation reference temperature range, analyzes the persistence of the temperature offset, and establishes a temperature offset change sequence; The heat diffusion anomaly determination submodule screens the node states with heat diffusion imbalance based on the temperature offset change sequence and the temperature fluctuation characteristics in a continuous time window, detects heat diffusion anomaly events, and obtains a heat diffusion anomaly trend indicator; The task migration mapping submodule extracts the nodes and tasks corresponding to the abnormal event based on the heat diffusion abnormal trend indicator, splits the corresponding tasks, and migrates the split tasks to multiple nodes according to the resource occupancy status of each node to generate heat-sensitive migration adjustment records.

Images