Backup task scheduling method based on load prediction and improved SAC algorithm

Abstract

The invention discloses a backup task scheduling method based on load prediction and an improved SAC algorithm, and relates to the technical field of data backup and intelligent scheduling. Aiming at the problems of response lagging, insufficient key task guarantee and low strategy exploration efficiency of a traditional scheduling method under a dynamic load, a task weight monitoring matrix is constructed, server performance, a network state and data importance are fused, and a task risk coefficient is calculated in combination with an exponential smoothing algorithm; carrying out sequence decomposition on a service load by adopting an Autoformer model, identifying a periodic mode through a weighted autocorrelation mechanism, and predicting a future load change trend; on this basis, an improved SAC scheduling framework is designed, a task weight guidance and temperature coefficient adaptive adjustment mechanism is introduced, and the task starting opportunity and execution duration are optimized in combination with an elastic time window. The invention provides an intelligent scheduling method with strong adaptive capability, which can autonomously optimize the task scheduling strategy according to the dynamic change of the system and effectively reduce the peak load of the server and the task backlog risk.

Description

Technical Field

[0001] This invention relates to the field of data backup and intelligent scheduling technology, and in particular to a backup task scheduling method based on load prediction and an improved Soft Actor-Critic (SAC) algorithm. Background Technology

[0002] In data center backup systems, the level of intelligence in scheduling strategies directly determines data security capabilities and system resource utilization efficiency. With the acceleration of enterprise digital transformation and the increasing dynamism of business workloads, backup tasks face complex operating environments such as server performance fluctuations, network congestion, and significant differences in data importance. Traditional static scheduling methods based on fixed time windows or priority queues are no longer sufficient to meet these demands.

[0003] In large-scale data center backup scenarios, scheduling decisions exhibit significant multidimensional coupling and long-cycle dependencies. Existing technologies fail to effectively integrate heterogeneous indicators such as server performance, network status, and data importance for risk modeling, and also lack the ability to accurately predict business load trends. Recent research has attempted to introduce sequence models (such as LSTM and Transformer) for load prediction and combine them with reinforcement learning to optimize scheduling strategies. However, these methods often focus on single-dimensional data analysis, failing to fully explore the periodic and trend components of the load, and their model update mechanisms lack closed-loop feedback, making it impossible to continuously optimize strategies based on actual execution results. Therefore, it is necessary to design an improved SAC closed-loop backup scheduling method based on a task weight monitoring matrix and Autoformer load prediction. By integrating multi-source state awareness, periodic pattern recognition, and adaptive strategy exploration, this method achieves dynamic collaborative optimization of task execution timing and resource consumption, thereby addressing the aforementioned technical problems. Summary of the Invention

[0004] In view of the shortcomings of the above-mentioned technologies, the purpose of this invention is to solve the problems of response lag under dynamic load, insufficient guarantee of critical tasks, and low efficiency of strategy exploration in the existing technologies.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The backup task scheduling method based on load prediction and improved SAC algorithm includes the following steps:

[0007] Step 1: Collect server performance, network bandwidth and latency information, as well as data importance level, and obtain the task risk coefficient through the exponential smoothing algorithm to construct the task weight monitoring matrix;

[0008] Step 2: Based on the task weight monitoring matrix and historical task patterns, use the Autoformer model to predict business load trends and generate dynamic load perception curves.

[0009] Step 3: Make scheduling decisions based on the improved SAC algorithm. Generate an adaptive scheduling strategy according to the task weight monitoring matrix and dynamic load perception curve. Dynamically adjust the execution timing and duration of each backup task to achieve coordinated optimization of load balancing and system stability.

[0010] Optionally, in step 1, the collected server performance status includes CPU utilization, memory usage, and I / O throughput, while network bandwidth and latency information includes link bandwidth utilization, end-to-end communication latency, and packet loss rate.

[0011] Data importance level is represented as This is used to distinguish the priority levels of different tasks, where 1 represents a low priority task and 5 represents a high priority task.

[0012] Let the data collection period be... The unit is seconds, and the collection duration is... Then the timestamp sequence is ,in ;

[0013] The above indicators are normalized at multiple time points and then integrated to form a system state history matrix. Its form is:

[0014]

[0015] in, This represents the number of acquisition cycles. Each row represents a 6-dimensional system state feature vector acquired at a certain moment, and each column represents a sequence of a certain type of index changing over time.

[0016] The data importance level From recovery time target and data sensitivity The decision was made jointly, and the specific calculation formula is as follows:

[0017]

[0018] in, The recovery time target required by the Service Level Agreement (SLA) represents the upper limit of the time that the system must complete to recover after a failure. This indicates the maximum allowed recovery time set by the system. This represents the maximum value of data sensitivity for all tasks in the system.

[0019] Furthermore, a single exponential smoothing method is used to weight and fuse the multi-source indicators, outputting the task risk coefficient for each dimension. The calculation formula is as follows:

[0020]

[0021] in, Indicates the first Dimensional metrics over time The observed values, As a smoothing factor, This represents the risk estimate for that dimension at the previous moment;

[0022] initial value Take the arithmetic mean of the observations for the first three periods of this dimension, that is:

[0023]

[0024] And smoothing factor The value is dynamically adjusted based on the volatility of the indicator, taking a larger value when the indicator fluctuates greatly and a smaller value when the indicator tends to stabilize.

[0025] By calculating the risk coefficients for all indicator dimensions sequentially, the final task risk coefficient vector is obtained. ;

[0026] Based on this, a task weight monitoring matrix is ​​constructed by combining the data importance levels. ,

[0027] in Let be the number of tasks to be scheduled in the k-th period, and let be the scheduling weight of the i-th task in the j-th dimension.

[0028]

[0029] That is, the weight of a task in a certain dimension is equal to the risk coefficient of the corresponding dimension multiplied by the priority of the task.

[0030] The proposed method maps server performance fluctuations, network congestion risks, and the importance of data services to the scheduling weight space, enabling highly sensitive and low-tolerance tasks to obtain stronger protection in resource competition. It effectively solves the problem of "high-priority tasks being delayed due to system overload" in traditional scheduling and significantly improves the SLA satisfaction rate of critical backup tasks.

[0031] Furthermore, the Autoformer model monitors the task weight matrix. and historical system load sequence As input,

[0032] in Indicates the first The system state vector at each moment corresponds to the six dimensions of CPU utilization, memory usage, I / O throughput, link bandwidth utilization, end-to-end communication latency, and packet loss rate, respectively.

[0033] The Autoformer model uses a sequence decomposition module to... Decomposed into trend term sequence and periodic term sequence The trend term reflects long-term patterns of change, while the periodic term reflects short-term recurring fluctuations.

[0034] In the encoder, the model uses an autocorrelation mechanism to identify historical periodic segments with similar load patterns:

[0035] Will According to the preset cycle length It is divided into multiple continuous subsequences, and each subsequence constitutes a periodic segment;

[0036] The period segment corresponding to the current time to be predicted is used as the query segment. The remaining historical fragments are used as candidate matching fragments. ;

[0037] For each task to be scheduled Add it to the task weight monitoring matrix The corresponding weight vector As the weight for similarity calculation of various system indicator dimensions;

[0038] Generate future data based on the fluctuation patterns of historical cycle segments with the highest weighted similarity. Periodic prediction components at each time step;

[0039] At the same time, for the trend term sequence Perform long-term trend modeling to generate corresponding trend prediction components;

[0040] The model overlays trend and cycle prediction components during the decoding phase to generate the future. Load forecast sequence at each time step ,right A weighted correction is applied to attenuate the load to the current measured load value. The attenuation factor is dynamically adjusted according to the anomaly confidence level, forming the final dynamic load perception curve. This is used to construct the state space of the SAC algorithm.

[0041] Compared to LSTM or ordinary Transformer, the proposed method significantly improves the accuracy and robustness of load prediction by decomposing trend and periodic terms, introducing a weighted autocorrelation mechanism guided by task weights, and combining anomaly perception correction to dynamically calibrate the prediction sequence. This provides reliable and forward-looking state input for scheduling strategies and effectively avoids task backlog.

[0042] Furthermore, within each scheduling cycle, the following scheduling decision steps are executed:

[0043] The obtained dynamic load sensing curve With task weight monitoring matrix The features are combined into a joint state through feature splicing. The data is input into a pre-trained scheduling strategy model.

[0044] The scheduling strategy model is built on the Soft Actor-Critic algorithm framework. Its output layer generates the mean and standard deviation of the action distribution of each task to be scheduled, and uses the reparameterization technique to sample specific actions from the distribution.

[0045] The model output includes the start time offset. and execution duration adjustment amount Both follow a normal distribution:

[0046]

[0047] The mean , This represents the optimal adjustment direction recommended by the model, expressed as standard deviation. , It reflects the degree of strategy exploration;

[0048] For each task, samples are independently generated based on its corresponding probability distribution. and ;

[0049] Based on the sampling results, the first The original start time of the task Adjusted to Original duration Adjusted to This creates a flexible time window and sends scheduling instructions to the execution module.

[0050] Furthermore, after generating the elastic time window, the system detects time overlap and shared resource conflicts between tasks; if the superimposed resource requirements exceed the limit, the window of the high-priority task is fixed, and the start time of the low-priority task is finely adjusted within the elastic range to alleviate the conflict.

[0051] To maximize scheduling benefits while maintaining policy exploration capabilities and preventing getting trapped in local optima, the optimization objective of the scheduling policy model is defined by the following objective function:

[0052]

[0053] in For policy entropy, For temperature coefficient, In strategy State-action joint distribution under the following conditions In the state The probability of taking an action. The comprehensive scheduling reward is calculated based on resource overload, task priority, and whether the deadline is met:

[0054]

[0055] in For the first The data importance level of each task This is a task completion status indicator. It takes a value of 1 if the task completion time does not exceed the deadline, and 0 otherwise. The peak resource utilization rate predicted after scheduling. To preset a safety threshold, This is the overload penalty coefficient;

[0056] Temperature coefficient With the total length of the flexible window Adaptive change, satisfying:

[0057]

[0058] in The initial temperature coefficient has a range of values. , >0 indicates adjustable gain. This is the baseline window length for the system under ideal load conditions, when Significantly greater than hour, Automatic decay suppresses excessive strategy exploration and improves scheduling stability in high-load scenarios.

[0059] This improved SAC framework achieves flexible time window adjustment through action distribution sampling and combines a temperature coefficient adaptive mechanism to achieve a balance between strategy exploration and utilization. Compared with standard DQN or fixed threshold scheduling, it can still maintain a task completion rate of over 90% under sudden high loads, and the peak resource utilization of the server is always lower than the preset safety threshold.

[0060] Preferably, the scheduling strategy model achieves stable training and continuous optimization during the offline training phase through experience replay and the target network mechanism;

[0061] The scheduling strategy model includes a policy network, two independent value Q networks, a state value V network, and a target state value V network.

[0062] The state-value V network is used to estimate the expected cumulative return of the current policy in a given state and provides a benchmark for calculating the objective value of the Q network.

[0063] The target state value V network slowly tracks the parameter changes of the state value V network through a soft update mechanism to ensure the stability of the target value;

[0064] During the training initialization phase, the parameters of the target state value V network are... Initialization and parameters of the V-network with state value Initialize to the same value;

[0065] After each scheduled task is completed, the actual completion time, peak resource usage, and abnormal events of each task are collected. The comprehensive scheduling reward is then calculated by combining the task priority and system load status. ,

[0066] and the state of the current scheduling cycle The scheduling actions performed The aforementioned reward and the status of the next scheduling cycle. The training samples are stored in the experience replay pool, which consists of a circular buffer.

[0067] The scheduling action Includes the start time offset for each task to be scheduled. and execution duration adjustment amount It is used to dynamically adjust the execution timing and resource usage window of tasks according to the system load;

[0068] When the number of accumulated samples in the experience replay pool reaches a preset threshold At that time, small batches of data are randomly sampled from them;

[0069] Based on the sampled data, using the state at the next time step Calculate the target Q-value using the output of the target state V network, and update the Q network parameters by minimizing the error between the current Q network output and the target Q-value.

[0070]

[0071] in Discount factor;

[0072] The policy network is updated using gradient ascent to maximize the weighted sum of the expected scheduling reward and the policy entropy.

[0073] The action value is estimated by taking the smaller value of the two Q-network outputs to suppress Q-value overestimation;

[0074] The parameters of the state-value network V are updated via gradient descent, making its output approximate the actual attainable value of the policy in the current state, while the parameters of the target state-value network V are softly updated via exponential moving average.

[0075]

[0076] in This is the soft update rate coefficient. and These represent the parameter sets of the state-value V network and the target state-value V network, respectively.

[0077] Throughout the training process, the target V network parameters always lag behind the state V network parameters to provide a stable estimate of the target value and suppress value function oscillations during training.

[0078] Compared to existing technologies, this solution has the following advantages:

[0079] By analyzing historical data on system load, task priority, and resource usage, a dynamic load perception curve and task weight monitoring matrix are generated. This comprehensively integrates temporal status and task attributes, providing accurate and quantifiable decision-making basis for scheduling. Based on this, resource conflict coordination fine-tuning is performed on the initial scheduling actions output by the strategy model. While ensuring the high-priority task window remains unchanged, the start time of low-priority tasks is offset with minimal disturbance, effectively alleviating congestion caused by shared resource overload during multi-task concurrency and improving resource utilization efficiency. Simultaneously, an anomaly perception and correction mechanism is introduced. When a deviation from the normal system state is detected, the predicted load is weighted and attenuated to the current measured value, avoiding erroneous scheduling due to inaccurate predictions during sudden failures. This approach responds to future load and priority changes while also considering current resource contention and anomaly risks, thereby optimizing task execution timing, rationally allocating computation and I / O resources, improving backup task completion rate and SLA satisfaction rate, and ensuring the stable and efficient operation of the disaster recovery system in complex environments. Attached Figure Description

[0080] Figure 1 This is an overall flowchart of a backup task scheduling method based on load prediction and an improved SAC algorithm.

[0081] Figure 2 To improve the SAC scheduling decision framework model structure diagram. Detailed Implementation

[0082] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0083] Example: A backup task scheduling method based on load prediction and an improved SAC algorithm, in Figure 1 The specific implementation methods for each stage of the overall process shown are as follows:

[0084] First, the system collects server performance, network status, and data importance information at fixed intervals. It then uses an exponential smoothing algorithm to calculate the task risk coefficient and combines it with priority to construct a task weight monitoring matrix, providing a dynamic perception basis for subsequent load prediction and scheduling decisions.

[0085] The collected server performance data includes CPU utilization, memory usage, and I / O throughput; network bandwidth and latency information includes link bandwidth utilization, end-to-end communication latency, and packet loss rate; data importance levels are represented as follows: This is used to distinguish the priority levels of different tasks, where 1 represents a low priority task and 5 represents a high priority task.

[0086] Let the data collection period be... The unit is seconds, and the collection duration is... Then the timestamp sequence is ,in ;

[0087] The above indicators are normalized at multiple time points and then integrated to form a system state history matrix. Its form is:

[0088]

[0089] in, This represents the number of acquisition cycles. Each row represents a 6-dimensional system state feature vector acquired at a certain moment, and each column represents a sequence of a certain type of index changing over time.

[0090] The data importance level From recovery time target and data sensitivity The decision was made jointly, and the specific calculation formula is as follows:

[0091]

[0092] in, The recovery time target required by the Service Level Agreement (SLA) represents the upper limit of the time that the system must complete to recover after a failure. This indicates the maximum allowed recovery time set by the system. This represents the maximum value of data sensitivity for all tasks in the system.

[0093] This design maps business continuity requirements and data value to a unified scheduling priority space, enabling highly sensitive and low-tolerance tasks to be better protected in resource competition.

[0094] Furthermore, a single exponential smoothing method is used to weight and fuse the multi-source indicators, outputting the task risk coefficient for each dimension. The calculation formula is as follows:

[0095]

[0096] in, Indicates the first Dimensional metrics over time The observed values, As a smoothing factor, This represents the risk estimate of this dimension at the previous moment, from which the risk vector is obtained. This reflects the current system's potential stress level across various resource dimensions.

[0097] initial value Take the arithmetic mean of the observations for the first three periods of this dimension, that is:

[0098]

[0099] And smoothing factor The system dynamically adjusts based on indicator volatility; when increased indicator volatility is detected... It automatically increases to 0.7 to enhance responsiveness, or decreases to 0.3 to filter out noise; this mechanism enables risk assessment to both quickly capture sudden loads and remain stable under steady conditions.

[0100] By calculating the risk coefficients for all indicator dimensions sequentially, the final task risk coefficient vector is obtained. Based on this, a task weight monitoring matrix is ​​constructed by combining the data importance levels. ,

[0101] in Let be the number of tasks to be scheduled in the k-th period, and let be the scheduling weight of the i-th task in the j-th dimension.

[0102]

[0103] That is, the weight of a task in a certain dimension is equal to the risk coefficient of the corresponding dimension multiplied by the priority of the task; this matrix gives high-priority tasks a greater scheduling weight in the dimension of high-risk resources, thereby prioritizing the execution of tasks that have a critical impact on system stability during the prediction and scheduling phases.

[0104] Furthermore, and historical system load sequence Input the Autoformer model, where Indicates the first The system state vector at each moment corresponds to the six dimensions of CPU utilization, memory usage, I / O throughput, link bandwidth utilization, end-to-end communication latency, and packet loss rate, respectively.

[0105] The Autoformer model first uses the sequence decomposition module to... Decomposed into trend terms and periodic terms This approach captures long-term trends and daily or weekly recurring patterns separately. This decomposition effectively alleviates the representational conflict caused by a single model simultaneously fitting trend and periodic components, thus improving prediction accuracy.

[0106] In the processing of periodic terms, According to the preset cycle length The sequence is divided into multiple continuous subsequences, each subsequence constituting a periodic segment. The periodic segment corresponding to the current time to be predicted is used as the query segment. The remaining historical fragments are used as candidate matching fragments. ;

[0107] For each task to be scheduled Add it to the task weight monitoring matrix The corresponding weight vector As a similarity calculation for various system indicator dimensions, historical periods that match high-weight dimensions receive more attention;

[0108] Ultimately, the model generates future data based on the fluctuation patterns of the historical periodic segments with the highest weighted similarity. Periodic forecast components at each time step; trend term Then, long-term trend modeling is performed to generate corresponding trend prediction components;

[0109] The model overlays trend and cycle prediction components during the decoding phase to generate the future. Load forecast sequence at each time step After anomaly detection correction, this sequence forms the final dynamic load sensing curve. .

[0110] Specifically, the system runs a lightweight anomaly detection module in parallel to determine in real time whether the current system state significantly deviates from the historical normal pattern. If an anomaly is detected, the initial load prediction sequence is weighted and corrected to decay towards the current measured value. Specifically:

[0111]

[0112] in To predict load, This is the current measured load. This is the decay factor, whose value decreases as the anomaly confidence decreases. The corrected sequence. That is, it is used as the input of the dynamic load perception curve to the scheduling strategy model.

[0113] Figure 1The process from initial state data collection to load prediction is now complete. This prediction mechanism shifts scheduling decisions from passive response to proactive avoidance, reserving execution windows before resource peaks arrive, effectively mitigating task backlog and service default risks.

[0114] During the scheduling decision-making phase, the obtained dynamic load perception curve will be... With task weight monitoring matrix The features are combined into a joint state through feature splicing. The input is fed into a policy network built on the SAC framework;

[0115] like Figure 2 As shown, the scheduling strategy model adopts an improved SAC architecture, which includes a policy network, two independent Q networks, a state value V network, and a target V network. Addressing the characteristics of sparse rewards and high action dimensionality in backup scheduling, this structure suppresses overestimation of value through the dual Q network, maintains training stability through soft updates of the target V network, and retains exploration capability by relying on policy entropy, thus avoiding suboptimal scheduling.

[0116] The network outputs the start time offset for each task. Adjustment amount with execution duration The probability distribution parameters (mean and standard deviation) of the given values ​​both follow a normal distribution:

[0117]

[0118] The mean , This represents the optimal adjustment direction recommended by the model, expressed as standard deviation. , It reflects the degree of strategy exploration;

[0119] For each task, samples are independently generated based on its corresponding probability distribution. and ;

[0120] Based on the sampling results, the first The original start time of the task Adjusted to Original duration Adjusted to This creates a flexible time window and sends scheduling instructions to the execution module.

[0121] To maximize scheduling benefits while maintaining policy exploration capabilities and preventing getting trapped in local optima, the optimization objective of the scheduling policy model is defined by the following objective function:

[0122]

[0123] in For policy entropy, For temperature coefficient, In strategy State-action joint distribution under the following conditions In the state The probability of taking an action. The comprehensive scheduling reward is calculated based on resource overload, task priority, and whether the deadline is met:

[0124]

[0125] in For the first The data importance level of each task This is a task completion status indicator. It takes a value of 1 if the task completion time does not exceed the deadline, and 0 otherwise. The peak resource utilization rate predicted after scheduling. To preset a safety threshold, This is the overload penalty coefficient;

[0126] Temperature coefficient With the total length of the flexible window Adaptive change, satisfying:

[0127]

[0128] in The initial temperature coefficient has a range of values. , >0 indicates adjustable gain. This is the baseline window length for the system under ideal load conditions, when Significantly greater than hour, Automatic decay suppresses excessive strategy exploration and improves scheduling stability in high-load scenarios;

[0129] Figure 2 The improved SAC architecture shown is the core of the aforementioned adaptive scheduling. It integrates task weights and load predictions as state inputs, enabling the strategy to proactively avoid resource overload while ensuring the SLA of critical tasks.

[0130] The scheduling strategy model achieves stable training and continuous optimization during the offline training phase through experience replay and the target network mechanism.

[0131] The state value V network is used to estimate the expected cumulative return of the current policy in a given state and provides a benchmark for the calculation of the target value of the Q network; the target state value V network slowly tracks the parameter changes of the state value V network through a soft update mechanism to ensure the stability of the target value.

[0132] During the training initialization phase, the parameters of the target state value V network are... Initialization and parameters of the V-network with state value Initialize to the same value;

[0133] Specifically, the scheduler first generates scheduling actions for each task to be scheduled based on the current system state. Includes the start time offset for each task to be scheduled. and execution duration adjustment amount It is used to dynamically adjust the execution timing and resource usage window of tasks based on system load.

[0134] Before issuing and executing scheduling actions, the scheduler further performs resource conflict coordination fine-tuning: it iterates through all pairs of tasks to be scheduled, checks whether their adjusted time windows overlap, and determines whether they share the same critical resource; if there is overlap and the superimposed resource demand exceeds a preset safety threshold, then based on the priority in the task weight monitoring matrix, the scheduling window of the high-priority task is fixed, and the start time of the low-priority task is offset with minimal disturbance within the allowable elastic adjustment range, until the resource superposition drops below the safety threshold or reaches the adjustment boundary. Only after fine-tuning is the scheduling scheme actually executed.

[0135] After each scheduled task is completed, the actual completion time, peak resource usage, and abnormal events of each task are collected. The comprehensive scheduling reward is then calculated by combining the task priority and system load status. and the state of the current scheduling cycle. The scheduling actions performed The aforementioned reward and the status of the next scheduling cycle. The training samples are stored in the experience replay pool, which consists of a circular buffer.

[0136] When the number of accumulated samples in the experience replay pool reaches a preset threshold At that time, small batches of data are randomly sampled from them for training;

[0137] Based on the sampled data, using the state at the next time step Calculate the target Q value using the output of the target state V network:

[0138]

[0139] in As a discount factor, The network minimizes the current output relative to the target. The error in the value is updated.

[0140] The policy network is updated using gradient ascent to maximize the weighted sum of expected scheduling reward and policy entropy, where the action value estimate is taken from the smaller of the two Q-network outputs to suppress Q-value overestimation.

[0141] The parameters of the state-value network V are updated via gradient descent, making its output approximate the actual attainable value of the policy in the current state, while the parameters of the target state-value network V are softly updated via exponential moving average.

[0142]

[0143] in This is the soft update rate coefficient. and These represent the parameter sets of the state-value V network and the target state-value V network, respectively.

[0144] Throughout the training process, the target V network parameters always lag behind the state V network parameters to provide a stable estimate of the target value, suppress value function oscillations during training, and ensure policy convergence.

[0145] Thus, this method forms a closed-loop mechanism of "perception-prediction-scheduling-feedback", which improves dynamic load response capability, effectively guarantees the SLA of high-priority tasks, and enhances the stability and adaptability of the strategy in complex environments.

[0146] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A backup task scheduling method based on load prediction and an improved SAC algorithm, characterized in that, The method includes the following steps: Step 1: Collect server performance, network bandwidth and latency information, as well as data importance level, and obtain the task risk coefficient through the exponential smoothing algorithm to construct the task weight monitoring matrix; Step 2: Based on the task weight monitoring matrix and historical task patterns, use the Autoformer model to predict business load trends and generate dynamic load perception curves. Step 3: Make scheduling decisions based on the improved SAC algorithm. Generate an adaptive scheduling strategy according to the task weight monitoring matrix and dynamic load perception curve. Dynamically adjust the execution timing and duration of each backup task to achieve coordinated optimization of load balancing and system stability.

2. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 1, characterized in that, In step 1, the collected server performance status includes CPU utilization, memory usage, and I / O throughput; network bandwidth and latency information includes link bandwidth utilization, end-to-end communication latency, and packet loss rate; data importance level is represented as follows: This is used to distinguish the priority levels of different tasks, where 1 represents a low priority task and 5 represents a high priority task. Let the data collection period be... The unit is seconds, and the collection duration is... Then the timestamp sequence is ,in The above indicators are normalized at multiple time points and then integrated to form a system state history matrix. Its form is: in, This represents the number of acquisition cycles. Each row represents a 6-dimensional system state feature vector acquired at a certain moment, and each column represents a sequence of a certain type of index changing over time. The importance level of the data From recovery time target and data sensitivity The decision was made jointly, and the specific calculation formula is as follows: in, The recovery time target required by the Service Level Agreement (SLA) represents the upper limit of the time that the system must complete to recover after a failure. This indicates the maximum allowed recovery time set by the system. This represents the maximum value of data sensitivity for all tasks in the system.

3. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 2, characterized in that, In step 1, a single exponential smoothing method is used to weight and fuse the multi-source indicators, outputting the task risk coefficient for each dimension. The calculation formula is as follows: in, Indicates the first Dimensional metrics over time The observed values, As a smoothing factor, This represents the risk estimate for this dimension at the previous moment; initial value. Take the arithmetic mean of the observations from the previous three periods for this dimension, that is: And smoothing factor The risk coefficient is dynamically adjusted based on the volatility of the indicators, taking larger values ​​when the indicators fluctuate significantly and smaller values ​​when the indicators tend to stabilize. By calculating the risk coefficients for all indicator dimensions sequentially, the final task risk coefficient vector is obtained. ; Based on this, a task weight monitoring matrix is ​​constructed by combining the data importance levels. ,in Let be the number of tasks to be scheduled in the k-th period, and let be the scheduling weight of the i-th task in the j-th dimension: That is, the weight of a task in a certain dimension is equal to the risk coefficient of the corresponding dimension multiplied by the priority of the task.

4. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 3, characterized in that, In step 2, the Autoformer model monitors the task weight matrix. and historical system load sequence As input, where Indicates the first The system state vector at each moment corresponds to the six dimensions of CPU utilization, memory usage, I / O throughput, link bandwidth utilization, end-to-end communication latency, and packet loss rate, respectively. The Autoformer model uses a sequence decomposition module to... Decomposed into trend term sequence and periodic term sequence The trend term reflects long-term patterns of change, while the periodic term reflects short-term recurring fluctuations. In the encoder, the model employs an autocorrelation mechanism to identify historical periodic segments with similar load patterns: According to the preset cycle length The time segment is divided into multiple continuous subsequences, each subsequence constituting a periodic segment; the periodic segment corresponding to the current time to be predicted is used as the query segment. The remaining historical fragments are used as candidate matching fragments. ; For each task to be scheduled Add it to the task weight monitoring matrix The corresponding weight vector As the weight for similarity calculation across various system indicator dimensions; based on the fluctuation patterns of the historical periodic segments with the highest weighted similarity, future... The periodic prediction component at each time step; simultaneously, the trend term sequence Perform long-term trend modeling to generate corresponding trend prediction components; The model overlays trend and cycle prediction components during the decoding phase to generate the future. Load forecast sequence at each time step ,right A weighted correction is applied to attenuate the load to the current measured load value. The attenuation factor is dynamically adjusted according to the anomaly confidence level, forming the final dynamic load perception curve. .

5. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 4, characterized in that, In step 3, the following scheduling decision steps are executed within each scheduling cycle: The currently obtained dynamic load perception curve With task weight monitoring matrix The features are combined into a joint state through feature splicing. The data is input into a pre-trained scheduling strategy model. The scheduling strategy model is built on the Soft Actor-Critic algorithm framework. Its output layer generates the mean and standard deviation of the action time distribution of each task to be scheduled, and uses the reparameterization technique to sample specific actions from the distribution. The model output includes the start time offset. and execution duration adjustment amount Both follow a normal distribution: The mean , This represents the optimal adjustment direction recommended by the model, and the standard deviation. , It reflects the degree of strategy exploration; for each task, it independently samples according to its corresponding probability distribution to generate specific [data / data]. and ; Based on the sampling results, the first The original start time of the task Adjusted to Original duration Adjusted to This creates a flexible time window and sends scheduling instructions to the execution module. To maximize scheduling benefits while maintaining policy exploration capabilities and preventing getting trapped in local optima, the optimization objective of the scheduling policy model is defined by the following objective function: in For policy entropy, For temperature coefficient, In strategy State-action joint distribution under the following conditions In the state The probability of taking an action. The comprehensive scheduling reward is calculated based on resource overload, task priority, and whether the deadline is met: in For the first The data importance level of each task This is a task completion status indicator. It takes a value of 1 if the task completion time does not exceed the deadline, and 0 otherwise. The peak resource utilization rate predicted after scheduling. To preset a safety threshold, This is the overload penalty coefficient; The temperature coefficient With the total length of the flexible window Adaptive change, satisfying: in The initial temperature coefficient has a range of values. , >0 indicates adjustable gain. This is the baseline window length for the system under ideal load conditions, when Significantly greater than hour, Automatic decay suppresses excessive strategy exploration and improves scheduling stability in high-load scenarios.

6. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 5, characterized in that, The scheduling strategy model achieves stable training and continuous optimization during the offline training phase through experience replay and the target network mechanism. The scheduling strategy model includes a strategy network, two independent value Q networks, a state value V network, and a target state value V network. The state value V network is used to estimate the expected cumulative return of the current strategy in a given state and to provide a benchmark for the target value calculation of the Q network. The target state value V network slowly tracks the parameter changes of the state value V network through a soft update mechanism to ensure the stability of the target value. During the training initialization phase, the target state value V is determined by the network parameters. Initialization and parameters of the V-network with state value Initialize to the same value; After each scheduled task is completed, the actual completion time, peak resource usage, and abnormal events of each task are collected. The comprehensive scheduling reward is then calculated by combining the task priority and system load status. and the state of the current scheduling cycle. The scheduling actions performed The aforementioned reward and the status of the next scheduling cycle. The training samples are stored in the experience replay pool, which consists of a circular buffer. The scheduling action Includes the start time offset for each task to be scheduled. and execution duration adjustment amount It is used to dynamically adjust the execution timing and resource usage window of tasks according to the system load; When the number of accumulated samples in the experience replay pool reaches a preset threshold At that time, small batches of data are randomly sampled from them; Based on the sampled data, using the state at the next time step Calculate the target Q-value using the output of the target state V network, and update the Q network parameters by minimizing the error between the current Q network output and the target Q-value. in The discount factor is used; the policy network is updated using gradient ascent to maximize the weighted sum of expected scheduling reward and policy entropy, where the action value estimate is taken from the smaller of the two Q network outputs to suppress Q-value overestimation. The parameters of the state-value network V are updated via gradient descent, making its output approximate the actual attainable value of the policy in the current state, while the parameters of the target state-value network V are softly updated via exponential moving average. in This is the soft update rate coefficient. and Let V represent the parameter sets of the state value V network and the target state value V network, respectively. Throughout the training process, the parameters of the target V network always lag behind the parameters of the state V network to provide a stable target value estimate and suppress value function oscillations during training.

7. The backup task scheduling method based on load prediction and improved SAC algorithm according to claim 5, characterized in that, After generating the flexible time windows for each task, the following steps are also included: The system detects whether the time windows of any two tasks overlap and determines whether they share the same critical resource. If there is an overlap and the superimposed resource demand in the shared resource dimension exceeds the preset safety threshold, the scheduling window of the high-priority task is kept unchanged according to the priority in the task weight monitoring matrix, and the start time of the low-priority task is fine-tuned within the elastic adjustment range until the resource conflict is alleviated or the adjustment boundary is reached.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the backup task scheduling method based on load prediction and improved SAC algorithm as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the backup task scheduling method based on load prediction and improved SAC algorithm as described in any one of claims 1-6.

Images