Adaptive balanced scheduling closed-loop optimization method and system for cross-region multi-available area operation and maintenance architecture

Abstract

The application provides an adaptive balanced scheduling closed-loop optimization method and system for a cross-region multi-available area operation and maintenance architecture, relates to the technical field of distributed system resource scheduling, and comprises the following steps: collecting multi-dimensional resource state data and a collection timestamp of each available area in real time, and obtaining historical multi-dimensional resource state data through persistence; fitting a natural evolution trend of local business according to the historical multi-dimensional resource state data; synchronously calculating a difference between a current scheduling decision time and the collection timestamp to obtain a collection decision time difference value; adding the collection decision time difference value and a pre-stored cross-domain link round-trip delay to obtain cross-domain perception lag time; and according to the cross-domain perception lag time, backtracking from the historical multi-dimensional resource state data to obtain a resource state snapshot at a historical real collection time. The application can effectively realize stable and balanced scheduling under the cross-region multi-available area operation and maintenance architecture.

Description

Technical Field

[0001] This invention relates to the field of distributed system resource scheduling technology, and in particular to an adaptive balanced scheduling closed-loop optimization method and system for cross-regional multi-availability zone operation and maintenance architecture. Background Technology

[0002] In a cross-regional multi-availability zone operation and maintenance architecture, in order to ensure the balance of resource utilization and the timeliness of service response among different availability zones, it is usually necessary to dynamically schedule stateless computing tasks or risk control verification requests among different availability zones.

[0003] However, most existing cross-regional scheduling technologies face the following technical shortcomings: there is a persistent perception lag between data acquisition and scheduling decision-making. Specifically, acquisition probes deployed in each availability zone collect multi-dimensional resource status data at fixed intervals and report it to the decision center; due to the time consumed in acquisition and the round-trip latency of cross-domain links, by the time the decision center generates scheduling instructions based on this data, the actual resource status of each availability zone has changed, and the scheduling decision is always based on an outdated picture that lags behind the actual status.

[0004] For example, because the decision center cannot know the real-time changes in the resource status of each availability zone during the perception lag time, the scheduling amount calculated based on the expired status often deviates from the actual number of tasks that need to be migrated: when the actual load has increased due to the natural evolution of local business, the scheduling amount based on historical snapshots is too small and insufficient to alleviate the resource pressure of the target availability zone; conversely, when the actual load has naturally decreased, the scheduling amount based on historical snapshots is too large, resulting in over-scheduling and resource waste; moreover, in continuous rounds of scheduling, the resource status changes caused by each round of scheduling also need to go through the perception lag before they can be perceived by the next round, forming a vicious cycle of scheduling, waiting, outdated decisions, and rescheduling, causing the load to fluctuate back and forth, and the system to be difficult to converge to a stable equilibrium. Summary of the Invention

[0005] This invention provides an adaptive balanced scheduling closed-loop optimization method and system for cross-regional multi-availability zone operation and maintenance architecture. It can effectively solve the technical problem in existing cross-regional scheduling technology that there is a perception delay between the data acquisition time and the scheduling decision time, which leads to inaccurate scheduling decisions and difficulty in the system converging to a stable equilibrium.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, an adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architectures, the method comprising: Real-time collection of multi-dimensional resource status data and collection timestamps from each availability zone, and persistence of historical multi-dimensional resource status data; fitting of the natural evolution trend of local services based on historical multi-dimensional resource status data; synchronous calculation of the difference between the current scheduling decision time and the collection timestamp to obtain the collection decision time difference; adding the collection decision time difference to the pre-stored cross-domain link round-trip delay to obtain the cross-domain perceived latency time. Based on the cross-domain sensing lag time, a snapshot of the resource status at the actual historical acquisition time is obtained by backtracking from historical multidimensional resource status data; time alignment compensation is performed based on the resource status snapshot and the natural evolution trend to obtain compensated multidimensional resource status data; based on the compensated multidimensional resource status data, the resource supply and demand difference between the source availability zone and the target availability zone is calculated to obtain the resource load potential gradient. Based on the pre-stored historical scheduling instruction sequence, the resource load fluctuation amplitude is statistically obtained; based on the resource load potential energy gradient and cross-domain sensing hysteresis time, the cross-domain sensing hysteresis compensation factor is calculated; based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude, the dynamic convergence weight is obtained. Based on the resource load potential gradient and dynamic convergence weight, the target migration amount and adaptive balancing scheduling instructions are calculated. According to the adaptive balancing scheduling instructions and the target migration amount, the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone are dynamically scheduled to the target availability zone.

[0007] Secondly, an adaptive balanced scheduling closed-loop optimization system for cross-regional, multi-availability zone operation and maintenance architectures includes: The data acquisition and hysteresis perception module is used to collect multi-dimensional resource status data and collection timestamps of each availability zone in real time, persist historical multi-dimensional resource status data, fit the natural evolution trend of local services based on the historical multi-dimensional resource status data, and calculate the difference in collection decision time and the pre-stored cross-domain link round-trip latency to obtain the cross-domain perception hysteresis time. The state compensation and gradient calculation module is used to backtrack the resource state snapshot at the actual historical acquisition time from the historical multidimensional resource state data based on the cross-domain sensing lag time, and to perform time alignment compensation on the resource state snapshot in combination with the natural evolution trend to obtain the compensated multidimensional resource state data, and to calculate the resource supply and demand difference between the source availability zone and the target availability zone to obtain the resource load potential gradient. The fluctuation assessment and weight generation module is used to calculate the resource load fluctuation amplitude based on the pre-stored historical scheduling instruction sequence, calculate the cross-domain sensing hysteresis compensation factor based on the resource load potential energy gradient and the cross-domain sensing hysteresis time, and generate dynamic convergence weights based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude. The migration decision and scheduling execution module is used to calculate the target migration amount based on the resource load potential gradient and the dynamic convergence weight, generate an adaptive balance scheduling instruction, and dynamically schedule the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone to the target availability zone based on the adaptive balance scheduling instruction and the target migration amount.

[0008] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0009] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0010] The above-described solution of the present invention has at least the following beneficial effects: By calculating the cross-domain perception lag time and performing time alignment compensation based on the resource status snapshot at the historical real collection time and the natural evolution trend of local business, the resource status data on which scheduling decisions are based is made closer to the real state at the current moment. This effectively solves the problem of scheduling deviation caused by perception lag and reasonably improves the accuracy of cross-regional scheduling decisions.

[0011] By using historical scheduling instruction sequences to statistically analyze resource load fluctuations and combining them with cross-domain sensing hysteresis compensation factors to generate dynamic convergence weights, adaptive convergence constraints are applied to the potential energy direction indicator. This suppresses the load swings caused by sensing hysteresis and scheduling feedback delay, enabling the system to quickly converge to a stable equilibrium state after multiple rounds of scheduling.

[0012] The method forms a closed-loop optimization link that integrates data acquisition, hysteresis perception, state compensation, gradient calculation, fluctuation assessment, dynamic weighting, and decision execution. Each link works in concert to incorporate multi-dimensional factors such as cross-domain link latency, data acquisition decision time difference, and historical scheduling disturbances into the decision framework. This effectively adapts to the characteristics of high data acquisition latency and complex link conditions in cross-regional multi-availability zone scenarios, and has strong engineering feasibility and scenario universality. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of an adaptive balanced scheduling closed-loop optimization system for cross-regional multi-availability zone operation and maintenance architecture provided by an embodiment of the present invention; Figure 3 This is a comparison chart of the resource status before and after compensation and the actual load deviation provided by an embodiment of the present invention; Figure 4 This is a response surface plot of the dynamic convergence weight as a function of hysteresis and fluctuation, provided by an embodiment of the present invention. Detailed Implementation

[0014] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0015] like Figure 1 As shown, embodiments of the present invention propose an adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architectures. The method includes the following steps: Step 1: Collect multi-dimensional resource status data and collection timestamps from each availability zone in real time, and persist the historical multi-dimensional resource status data; fit the natural evolution trend of local services based on the historical multi-dimensional resource status data; synchronously calculate the difference between the current scheduling decision time and the collection timestamp to obtain the collection decision time difference; add the collection decision time difference to the pre-stored cross-domain link round-trip delay to obtain the cross-domain perceived latency time. Step 2: Based on the cross-domain sensing lag time, retrieve the resource status snapshot at the actual historical acquisition time from the historical multidimensional resource status data; perform time alignment compensation based on the resource status snapshot and the natural evolution trend to obtain the compensated multidimensional resource status data; calculate the resource supply and demand difference between the source availability zone and the target availability zone based on the compensated multidimensional resource status data to obtain the resource load potential gradient. Step 3: Based on the pre-stored historical scheduling instruction sequence, the resource load fluctuation amplitude is statistically obtained; based on the resource load potential gradient and cross-domain sensing hysteresis time, the cross-domain sensing hysteresis compensation factor is calculated; based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude, the dynamic convergence weight is obtained. Step 4: Based on the resource load potential gradient and dynamic convergence weight, calculate the target migration amount and adaptive balancing scheduling instructions; according to the adaptive balancing scheduling instructions and the target migration amount, dynamically schedule the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone to the target availability zone.

[0016] In this embodiment of the invention, the cross-domain perception lag time is obtained by superimposing the difference in acquisition decision time with the round-trip delay of the cross-domain link, which realizes the accurate quantification of the lag of the entire link of data acquisition and network transmission, and provides a reliable time reference for subsequent state compensation. Based on the cross-domain perception lag time, the resource state snapshot of the historical real acquisition time is traced back, and time alignment compensation is performed in combination with the natural evolution trend of local services, so that the resource state on which the scheduling decision is based is close to the real instantaneous state at the current moment, which effectively overcomes the defects of the prior art in making decisions based on expired images and reasonably improves the calculation accuracy of the scheduling quantity.

[0017] By statistically analyzing historical scheduling instruction sequences to obtain the fluctuation range of resource load, and integrating cross-domain perception hysteresis compensation factors to generate dynamic convergence weights, it can adaptively perceive the degree of load oscillation caused by historical scheduling. When the system fluctuates violently, it automatically converges the scheduling intensity, and appropriately releases the scheduling capacity when the system tends to be stable. This breaks the vicious cycle of scheduling, waiting, outdated decisions, and rescheduling in existing technologies, and accelerates the convergence of the load of each availability zone to a stable equilibrium.

[0018] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Based on the collection probes deployed in each availability zone, multi-dimensional resource status data is acquired in real time. Each piece of multi-dimensional resource status data is marked with a corresponding collection timestamp, forming multi-dimensional resource status data marked with collection timestamps. Specifically, this includes: Within each availability zone of the cross-regional multi-availability zone operation and maintenance architecture, an independently running and lightweight collection probe process is pre-deployed. This collection probe process is embedded in the operating system layer of each computing node as a kernel module. In the metric acquisition path, the collection probe process obtains CPU utilization and memory usage by directly reading the proc file system exposed by the operating system kernel. The proc file system is a memory-based virtual file system provided by the Linux kernel. This virtual file system maps the process control block and system running status data structure in the kernel to user-readable text information in real time in the form of a file directory tree. The collection probe process calculates CPU utilization by reading the utime and stime fields in the virtual status files corresponding to each process under the proc file system, and obtains memory usage by reading the resident memory size field in the memory mapping information file corresponding to each process.

[0019] This data acquisition probe process obtains network throughput by reading network device statistics interfaces in the sysfs virtual file system. The sysfs virtual file system is a hierarchical virtual file system provided by the Linux kernel that exports kernel device models and driver information to user space. This virtual file system mounts the transmit and receive byte counters and transmit and receive packet counters of each network device in the sys directory tree in the form of device attribute files. The data acquisition probe process calculates network throughput by reading the receive byte count and transmit byte count in the statistics attribute files of the corresponding network interface devices under the sys file system. The data acquisition probe process obtains the task queue depth by reading the queue length counter maintained by the task scheduler in shared memory. The task scheduler is an intermediate layer component deployed within each availability zone, responsible for receiving upstream task requests and distributing tasks to each compute node for execution according to a preset strategy. The task scheduler maintains an atomically updated queue length counter in the shared memory area, which records the number of tasks currently waiting to be assigned and processed in real time.

[0020] The four types of real-time operating indicators obtained above together constitute node-level multidimensional resource status data. The specific working process of the acquisition probe process is as follows: During the initialization phase, the acquisition probe process requests a fixed-length circular buffer queue in memory using a first-in-first-out strategy. This circular buffer queue is used to temporarily store the multidimensional resource status data from the most recent rounds. Whenever the preset fixed acquisition period timer expires and triggers an interrupt signal, the acquisition probe process responds to the interrupt signal and initiates a full polling operation. Based on the pre-maintained list of online computing nodes in the availability zone, it sequentially initiates remote retrieval requests for the above four types of indicators to each computing node in the list, and stores the multidimensional resource status data from the same round. The real-time operating metrics of all computing nodes retrieved within the current round are aggregated and merged according to node identifiers to form multidimensional resource status data for the current round, representing the overall resource status of the availability zone. After the aggregation and merging are completed, the acquisition probe process immediately calls the high-precision time acquisition interface to read the current time based on the local system clock of the host machine where the acquisition probe is located. This current time is used as the acquisition timestamp that is uniformly bound to the multidimensional resource status data acquired this time. The acquisition timestamp and the multidimensional resource status data are encapsulated into a data record in the form of key-value pairs and pushed into the circular buffer queue, thereby forming multidimensional resource status data marked with acquisition timestamps that are precisely bound to the acquisition time.

[0021] Step 1.2 involves persistently storing the multidimensional resource status data marked with collection timestamps in a time-series manner to obtain historical multidimensional resource status data. Specifically, this includes: after completing timestamp marking, the collection probes in each availability zone report the multidimensional resource status data marked with collection timestamps to a centrally deployed decision center in real time via the availability zone's internal LAN and cross-domain backbone network. The decision center is deployed using a primary-backup active-active architecture, comprising a data storage layer, a status calculation engine, and a scheduling decision engine. These three components interact via an internal high-speed message bus. The data storage layer incorporates a time-series database engine and... The key-value storage engine and time-series database engine are used to persistently store multi-dimensional resource status data reported by each availability zone in a time-series manner. The key-value storage engine is also used to store historical scheduling instruction sequences and network link quality information tables. The status calculation engine has built-in trend fitting, status compensation, and gradient calculation modules, which are responsible for performing trend fitting of resource status, time alignment compensation, and resource load potential gradient calculation, respectively. The scheduling decision engine has built-in weight calculation and instruction generation modules, which are responsible for calculating dynamic convergence weights and generating adaptive balanced scheduling instructions, respectively, and issuing scheduling instructions through the persistent network connection established with the task scheduler of each availability zone.

[0022] Regarding data persistence, after receiving multidimensional resource status data marked with collection timestamps from each availability zone, the data storage layer performs an ordered write operation in the time-series database engine using the collection timestamp as the index key, appending each data record to the persistent storage medium in chronological order. After continuous accumulation through multiple collection cycles, a continuous time series record with the collection timestamp as a strictly monotonically increasing index is formed in the time-series database engine of the data storage layer. This continuous time series record is the historical multidimensional resource status data. At any subsequent moment, the resource status snapshot at the corresponding time can be retrieved and extracted from it through the range query interface of the time-series database engine according to the specified time position.

[0023] Step 1.3: Based on historical multidimensional resource status data, extract the trend characteristics of local business load changes over time, and fit the natural evolution trend of local business. Specifically, this includes: using the trend fitting module in the status calculation engine to read the time series continuously recorded in the recent collection periods of historical multidimensional resource status data from the time series database engine in the data storage layer; performing pattern recognition on the change trajectory of local business load indicators as the collection timestamp progresses in the time series; extracting the alternating peaks and troughs of local business load over time with a daily cycle and the directional characteristics of long-term growth or decline with a weekly cycle; and combining the above-mentioned periodic fluctuation patterns and long-term directional characteristics to form trend features.

[0024] Based on the extracted trend features, a linear regression fitting method based on a sliding time window is used to fit the trend features to a curve. The fitting process follows the following fitting calculation formula: [The formula is missing from the original text, so the translation ends here.] Within a time window consisting of several collection cycles, let the first... Each collection timestamp is The corresponding local business load metric observation value is ,in The value ranges from 1 to integers and The time window length parameter is pre-configured; then the fitted natural evolution trend function is obtained. satisfy ,in Let be the trend slope parameter, and its expression is as follows: ; The baseline intercept parameter and The above summation symbols All expressed their opinion From 1 to Perform iterative summation to obtain the trend slope parameter. molecular part The product of window length and total time-load covariance is subtracted from the product of total time and total load. This difference, after normalizing the denominator, yields the optimal estimate of the linear correlation coefficient between time and load in the least squares sense. The denominator of the linear regression slope calculation reflects the dispersion of the time independent variable within the window; The positive and negative signs indicate the long-term growth trend and long-term decay trend of the load relative to the time axis, respectively. The absolute value of the baseline intercept parameter indicates the rate of change in the trend. molecular part This represents the cumulative total load minus the product of the trend slope parameter and the cumulative total over time, with the difference calculated over a window of length. After normalization, the baseline load level, i.e., the intercept value, corresponding to the start time of the current sliding time window is obtained. The value reflects the baseline load level, i.e., the intercept value, corresponding to the start time of the current sliding time window; the natural evolution trend function It is an evolution curve that conforms to the historical change pattern and can be extrapolated along the time axis, used to describe the natural change trend of local business load over time under the condition of no external scheduling intervention.

[0025] Step 1.4: Calculate the time difference between the current scheduling decision time and the data collection timestamp to obtain the data collection decision time difference value. Specifically, this includes: at the instant a new round of scheduling decision process is started, recording the current scheduling decision time based on the system clock of the host machine where the decision center is located. The current scheduling decision moment Accurate to the millisecond level; then, extract the associated collection timestamp from the currently used multidimensional resource status data marked with collection timestamps. The current scheduling decision time With collection timestamp The difference is calculated by performing a subtraction operation on the same monotonically increasing time coordinate system, and the resulting difference is denoted as the data acquisition and decision-making time difference. Its calculation follows the formula The unit of the calculated result is milliseconds (ms), representing the time difference between data acquisition and decision-making. It quantifies the time elapsed from the moment the acquisition probe completes data acquisition to the moment when scheduling decisions are generated based on that data. Specifically, it covers the buffering time within the acquisition probe, the data transmission time across the network, and the turnaround time for writing and reading data in the data storage layer of the decision center. In other words, it is the end-to-end time consumed from the completion of data acquisition to the entry of data into the decision-making process.

[0026] Step 1.5: Based on the pre-stored network link quality information, extract the round-trip time between the source availability zone and the target availability zone to obtain the pre-stored cross-domain link round-trip time. Specifically, this includes: a pre-stored network link quality information table is maintained in the key-value storage engine of the data storage layer. This table is obtained by sending bidirectional probe data packets to the cross-domain network links between each availability zone at a fixed probe period and measuring their round-trip times. The table stores the source availability zone identifier and the target availability zone identifier as the key, and the average round-trip time value obtained from multiple measurements within the probe period as the value. After determining the source availability zone and the target availability zone involved in this round of scheduling, a query key is constructed using the source availability zone identifier and the target availability zone identifier. A key-value query operation is performed from the pre-stored network link quality information table to extract the round-trip time value between the source availability zone and the target availability zone. This value is recorded as the pre-stored cross-domain link round-trip time. Its unit is milliseconds (ms); this cross-domain link round-trip delay This reflects the round-trip network transmission time from the decision center through the network exit on the source availability zone side, through the cross-domain backbone network to the task scheduler in the target availability zone, and then back to confirm the response.

[0027] Step 1.6: Add the acquisition-decision time difference to the pre-stored cross-domain link round-trip delay to obtain the cross-domain sensing hysteresis time. Specifically, this includes: adding the acquisition-decision time difference... Round-trip latency with the pre-stored cross-domain link The result of summing numerical values ​​under the same dimension is denoted as the cross-domain sensing hysteresis time. Its calculation follows the formula The unit of the calculated result is milliseconds (ms); this is the cross-domain sensing hysteresis time. The overall time delay encompasses the buffering time within the acquisition probe, the network transmission time for multi-dimensional resource status data to be reported from the acquisition probe to the decision center, the data writing and reading turnaround time within the decision center, and the round-trip network transmission time for scheduling instructions to be issued from the decision center to the target availability zone. These factors collectively constitute the end-to-end time delay from the completion of raw data acquisition to the actual effective time of the scheduling instructions in the target availability zone; this cross-domain sensing lag time... It accurately quantifies the lag in the scheduling decision-making process's perception of the actual resource status, providing a unified and accurate time reference parameter for time alignment compensation of resource status in subsequent steps.

[0028] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1, based on the cross-domain sensing hysteresis time, determines the historical backtracking time offset, specifically including: calculating the cross-domain sensing hysteresis time... Then, the cross-domain sensing hysteresis time This is directly defined as the historical backtracking time offset. This historical backtracking time offset represents the time span from the current scheduling decision moment back along the historical timeline, and its dimensions are the same as the cross-domain sensing hysteresis time. Maintaining consistency means being on the order of milliseconds; the physical meaning of the historical backtracking time offset is that, due to the end-to-end time delay introduced by data collection, network transmission, and decision processing in the cross-regional multi-availability zone operation and maintenance architecture, the resource status data on which the current scheduling decision is based actually reflects the cross-domain perceived latency time. The previous availability zone's actual resource status requires tracing back along the time axis from the current scheduling decision point to the historical time indicated by the offset, in order to locate the historical data collection time that strictly corresponds to the resource status data held by the current decision-making stage.

[0029] In determining this historical backtracking time offset, the current scheduling decision time recorded above will be used. As the starting anchor point for time reversal, cross-domain perception of lag time will be used. As the backtracking step size, the two together constitute the two basic parameters of the time backtracking operation. The historical backtracking time offset determined by this will serve as the key input parameter for subsequent steps to perform time positioning retrieval in historical multidimensional resource status data, providing an accurate time reference for subsequent time alignment compensation operations.

[0030] Step 2.2: Based on the historical backtracking time offset and the historical multidimensional resource status data, locate and extract the multidimensional resource status at the corresponding time to obtain a snapshot of the resource status at the actual historical acquisition time. Specifically, this includes: using the recorded current scheduling decision time... Using the time reference point, the current scheduling decision time will be... Subtract the determined historical backtracking time offset The difference obtained is the actual historical data collection time. The actual historical moment of collection The precise point in time at which the acquisition probe process actually completed the acquisition of multi-dimensional resource status data and bound the acquisition timestamp satisfies the following relationship: Based on this, the actual historical data collection time was used. As the query key for the range query interface of the time-series database engine, it performs precise time-based retrieval operations within the continuous time-series records composed of historical multi-dimensional resource status data in the data storage layer. This retrieval operation utilizes the ordered index structure built by the time-series database engine using the collection timestamp as the index key, and quickly locates the corresponding data using binary search or B+ tree index traversal. The most recent multidimensional resource status data record on the time axis.

[0031] Since the data storage layer in step 1.2 uses the collection timestamp as the index key for ordered writing and persistent storage, and the collection probe process marks the collected data records with high-precision timestamps in each collection cycle, it can ensure that the records returned by this retrieval operation are consistent with the records in the time dimension. Strict alignment; the extracted multidimensional resource status data record is a snapshot of the resource status at the actual historical acquisition moment, and this resource status snapshot completely contains the actual historical acquisition moment. The multi-dimensional resource status information of the source and target availability zones covers the specific values ​​of four types of indicators at that moment: CPU utilization, memory usage, network throughput, and task queue depth. This provides real baseline status data for forward compensation along the time axis in subsequent steps.

[0032] Step 2.3: Based on the resource state snapshot and the natural evolution trend, perform forward compensation on the resource state snapshot along the time axis to obtain compensated multidimensional resource state data. Specifically, this includes: acquiring the historical real collection time at the location... After taking a resource status snapshot, the status values ​​of each dimension related to local business load metrics in the resource status snapshot are used as the baseline input for compensation calculation. Similar indicators, based on their historical data collection time. The specific value is denoted as ,in This indicates that the value is derived from a historical snapshot of the resource status. Identify the indicator category and The values ​​are traversed through four categories of metrics: CPU utilization, memory usage, network throughput, and task queue depth.

[0033] Based on this, the natural evolution trend function generated by the trend fitting module in step 1.3 within the same sliding time window is invoked. The natural evolution trend function It describes the natural variation of local business load over time under conditions of no external scheduling intervention, and further, it analyzes the load along the timeline from historical data collection points. Proceed forward to the current scheduling decision time. Calculate the time interval for advancement. Internal natural evolution trend function The increment of the function value, i.e. This increment This represents the amount of load change that should occur in the local business load from the historical data collection time to the current scheduling decision time, under the condition of no external scheduling intervention, according to the natural evolution law.

[0034] Further, this load change Calculations were performed separately for each indicator dimension, specifically for the [specific indicator]. Indicators are introduced with their corresponding trend conversion coefficients. The trend conversion factor Used to map load changes to state changes in this metric dimension. The calibration process is as follows: extract the most recent data from historical multidimensional resource status data. Within the first collection cycle, the first Observation sequence of class indicators Corresponding local service load observation sequence ,in A typical value is 100 acquisition cycles. The number in parentheses indicates the acquisition cycle number and its value ranges from 1 to 100. positive integers, Indicates the first Class indicators in the first Observations from each collection period, Indicates the first The local service load observation value for the first collection period; calculate the first... Pearson linear correlation coefficient between the observed series of class indicators and the observed series of local business load The formula for its calculation is: ; in For the first The arithmetic mean of the observation series of class indicators. This is the arithmetic mean of the local service load observation sequence. Indicates to From 1 to Perform iterative summation; obtain the trend conversion factor. ,Right now The value is the Pearson linear correlation coefficient. The absolute value; when Less than the preset weak correlation threshold and When the typical value is 0.3, Forced to be set to 0, indicating that the first... There is no significant linear correlation between this metric dimension and local business load; therefore, this metric dimension will not participate in the subsequent load change calculation process. After the above calibration is completed, the changes in each dimension after calculation will be... The status values ​​of the corresponding indicator dimensions are respectively overlaid into the historical resource status snapshot. Above, the compensated state values ​​for each dimension are obtained. ;in This indicates that the value is the compensated state value.

[0035] The compensated state values ​​of the above four types of indicators together constitute the compensated multidimensional resource state data. This compensated multidimensional resource state data reflects the historical data collected at the actual time point, assuming no external scheduling intervention. Up to the current scheduling decision time The theoretical state value that the resource status should reach according to the laws of natural evolution during this period; to intuitively verify the effect of time alignment compensation on the correction of perception lag, see [reference needed]. Figure 3 The graph shows a comparison of resource status before and after compensation with the actual load. The horizontal axis represents three resource indicators from left to right: CPU utilization, memory usage, and task queue depth. The vertical axis represents resource status values ​​(where CPU utilization and memory usage are normalized values ​​from 0 to 1, and task queue depth is in units). The light red bars on the left represent the historical data collection times. The snapshot of the resource status shows the original values. The light blue bars in the middle represent the multi-dimensional resource status data after compensation for natural evolution trends, and the dark green line represents the actual instantaneous load measured in the same time period.

[0036] Taking the scheduling rounds from location A to location B selected in this experiment as an example, the uncompensated original snapshot showed a CPU utilization of 72.0%, a memory utilization of 68.0%, and a task queue depth of 820. This deviated from the actual measured load (CPU 73.2% ± 0.3%, memory 69.1% ± 0.2%, queue depth 831 ± 5) by an average of 8.5% to 8.9%, representing typical lag distortion. After forward compensation, the CPU utilization increased to 73.9%, memory utilization increased to 69.9%, and the queue depth was adjusted to 834. The deviation from the measured values ​​narrowed to 0.8% to 1.0%, with an average deviation of only about 0.9%. This effectively eliminated the risk of misjudgment of the state caused by cross-domain perception lag, providing an accurate input benchmark for subsequent calculations of resource supply and demand differences. This forward compensation operation eliminated the time lag caused by cross-domain perception. The resulting lag in resource status perception makes it difficult to make timely scheduling decisions. The resource status data used is aligned with the actual resource status of the availability zone in the time dimension.

[0037] Step 2.4: Based on the compensated multidimensional resource status data, determine the resource supply status of the source availability zone and the resource demand status of the target availability zone. Specifically, this includes: after obtaining the compensated multidimensional resource status data, extracting the corresponding multidimensional resource status subsets for the source availability zone and the target availability zone from the compensated multidimensional resource status data according to the availability zone identifier; and extracting the compensated CPU utilization rate from the multidimensional resource status subset corresponding to the source availability zone. Memory usage after compensation and the compensated task queue depth Three types of indicators, among which This indicates that the indicator belongs to the source availability zone. The above three types of indicators are comprehensively calculated according to the preset supply capacity assessment weights to obtain the resource supply status of the source availability zone. The calculation method is as follows: ; in Assess the weighting of the CPU's surplus capacity. Weights for evaluating memory surplus capacity. The evaluation weights for the idle ratio of the task queue are determined by the following three factors: All values ​​are positive real numbers between 0 and 1. In this embodiment, the typical value for the supply capacity assessment weight is... , , The above weights can be adjusted and configured according to the contribution of various resources to the system service capabilities in actual business scenarios; The maximum capacity of the queue length counter maintained by the task scheduler in shared memory, i.e., the upper limit of the task queue depth; this resource provisioning status. The surplus resource capacity that the source availability zone can currently provide is quantified in scalar numerical form, with values ​​ranging from 0 to 1. The closer the value is to 1, the more abundant the surplus resources that the source availability zone can provide. At the same time, for the multi-dimensional resource state subset corresponding to the target availability zone, the compensated central processing unit utilization is extracted from it. Memory usage after compensation and the compensated task queue depth The three categories of indicators are combined and calculated according to preset urgency assessment weights to obtain the resource demand status of the target availability zone. The calculation formula is as follows: ; In the formula As a weight for evaluating the load pressure on the central processing unit, As a weight for evaluating memory load stress, The three factors satisfy the following weights for evaluating the congestion level of the task queue: All values ​​are positive real numbers between 0 and 1. In this embodiment, the typical value for the urgency assessment weight is... , , The above weights can be adjusted and configured according to the impact of various resource bottlenecks on system performance in actual business scenarios; The meaning is the same as before, namely the upper limit of the task queue depth; this resource requirement status. The resource gap that the target availability zone currently needs to meet is quantified in scalar numerical form. Its value range is also a real number between 0 and 1. The closer the value is to 1, the more urgent the resource demand of the target availability zone is.

[0038] It should be noted that the network throughput metric collected in step 1.1 does not directly participate in the scalarized supply and demand status assessment in the resource supply and demand status calculation of this step; the network throughput metric is used to assist in determining whether there is a bandwidth bottleneck in the cross-domain link: when the network throughput in either direction between the source availability zone and the target availability zone exceeds the preset bandwidth bottleneck threshold (typically 80% of the link's nominal bandwidth), the scheduling decision engine will adjust the calculated target migration volume in subsequent step 4.2. Apply additional network constraint factors , The typical value is 0.5, that is To avoid further exacerbating network bandwidth pressure during cross-domain task migration; resource supply status of the source availability zone. and the resource demand status of the target availability zone Both are calculated based on the same compensated multidimensional resource state data. They maintain strict consistency in terms of time dimension and data basis, providing a unified and comparable quantitative basis for calculating the resource load potential gradient in subsequent steps.

[0039] Step 2.5: Based on the resource supply status of the source availability zone and the resource demand status of the target availability zone, calculate the resource supply and demand difference between the resource supply status and the resource demand status to obtain the resource load potential gradient. Specifically, this includes: determining the resource supply status of the source availability zone... Resource demand status of the target availability zone Next, the resource requirement status of the target availability zone will be determined. Resource supply status of the source availability zone The difference obtained by performing subtraction under the same units of measurement is the resource supply and demand difference, which is then directly determined as the resource load potential energy gradient. The calculation method is as follows: ; The resource load potential gradient The physical meaning of this is that it quantifies the relative difference in resource pressure between the source and target availability zones from the perspective of resource supply and demand balance, and the positive or negative sign of its value indicates the expected direction of resource flow; specifically, when Right now This indicates that the resource demand in the target availability zone is more urgent than the resource surplus in the source availability zone, creating a potential driving force for cross-domain task scheduling from the source availability zone to the target availability zone. The larger the absolute value, the stronger the driving force; when Right now When the resource demand of the target availability zone does not exceed the resource surplus of the source availability zone, and there is no significant resource transfer potential demand in the target availability zone relative to the source availability zone, the current round does not trigger cross-domain scheduling operations from the source availability zone to the target availability zone.

[0040] Ultimately, this resource load potential gradient The absolute value of this parameter will serve as the core input parameter for the adaptive balanced scheduling instruction generation by the scheduling decision engine in step 3. The magnitude of this value directly determines the basis for selecting the dynamic convergence weight in subsequent scheduling decisions, i.e. The larger the value, the greater the dynamic convergence weight and the larger the scale of cross-domain task scheduling determined in this round of scheduling decisions. The smaller the value, the smaller the dynamic convergence weight and the smaller the scale of cross-domain task scheduling determined in this round of scheduling decisions, thereby achieving adaptive matching between the cross-domain scheduling intensity and the degree of resource supply and demand difference between availability zones.

[0041] In a preferred embodiment of the present invention, step 3 above may include:

[0042] Step 3.1: Based on the pre-stored historical scheduling instruction sequence, extract the scheduling action type and corresponding scheduling quantity from the historical scheduling instruction sequence at each historical scheduling moment, and statistically analyze the resource load change range of the source availability zone and the target availability zone before and after scheduling to obtain the resource load fluctuation amplitude. Specifically, this includes: before the scheduling decision engine executes the current round of cross-domain scheduling decision, retrieving the pre-stored historical scheduling instruction sequence from the data storage layer. This historical scheduling instruction sequence is composed of scheduling instruction records generated and executed in the past several scheduling cycles arranged in chronological order. Each historical scheduling instruction record contains the historical scheduling moment corresponding to that round of scheduling. Scheduling action types and scheduling quantities ,in The sequence number of the historical scheduling round and The value ranges from 1 to positive integers, This represents the total number of historical scheduling rounds contained in the historical scheduling instruction sequence. The typical value is 50 rounds. This value can achieve a reasonable balance between statistical stability and timeliness, ensuring that there are enough historical samples to support the credibility of the fluctuation range statistics, while avoiding the loss of reference value of historical records that are too far back due to changes in the resource topology of the availability zone.

[0043] For each historical scheduling instruction record, extract the historical scheduling time. Resource load status of the source availability zone before executing scheduling actions Resource load status after executing scheduling actions and the resource load status of the target availability zone before the scheduling action is performed. Resource load status after executing scheduling actions The resource load status mentioned above is based on the overall resource utilization rate of the availability zone at the corresponding time. The comprehensive resource utilization rate is measured as a real number between 0 and 1. Defined as a weighted sum of the status values ​​of four metrics for the availability zone at a given time: CPU utilization, memory usage, network throughput, and task queue depth. The formula is as follows: ; in, For indicator category identification, These correspond to CPU utilization, memory usage, network throughput, and task queue depth, respectively. For the first The normalized state value of the index at the corresponding time. ; For the first The weight coefficients corresponding to the class indicators satisfy the following conditions: and Weighting coefficient The values ​​are pre-calibrated through historical statistical analysis of the impact of each indicator on the overall load of the availability zone. The calibration method is as follows: extract the observation value sequence of each indicator and the overall load sequence of the availability zone from the historical operation data, calculate the correlation strength between each indicator and the overall load, and use the normalized result of the correlation strength as the weight coefficient of each indicator; based on the above definition, calculate the change in resource load before and after scheduling for the source availability zone. Calculate the change in resource load before and after scheduling for the target availability zone. ,Will and The larger value in is used as the first... Single-round resource load fluctuation corresponding to historical scheduling ,Right now The physical meaning of this maximum value operation is that cross-domain scheduling operations will cause resource status impacts on both the source and target availability zones. Using the larger fluctuation of the two as a measure of the degree of fluctuation in this round of scheduling can more conservatively reflect the maximum load oscillation caused by the scheduling operation.

[0044] Furthermore, on The corresponding historical scheduling round Statistical analysis was performed on the resource load fluctuations of each single round, and the standard deviation was calculated as the resource load fluctuation amplitude. The calculation method is as follows ;in This represents the arithmetic mean of the resource load fluctuations in each of the K rounds of historical scheduling; the magnitude of this resource load fluctuation... From the perspective of historical experience, the impact and fluctuation range of cross-domain scheduling operations on the resource load status of the source and target availability zones are quantified. The larger the value, the more severe the resource load oscillation caused by the cross-domain scheduling operation. In subsequent steps, the scheduling convergence speed needs to be adjusted accordingly.

[0045] Step 3.2: Based on the resource load potential energy gradient and the cross-domain sensing hysteresis time, calculate the degree of attenuation of the resource load potential energy gradient by the cross-domain sensing hysteresis time, and obtain the cross-domain sensing hysteresis compensation factor. Specifically, this includes: calculating the resource load potential energy gradient... Furthermore, step 1.6 calculates the cross-domain sensing hysteresis time. Next, the cross-domain sensing lag time will be... As input parameters, calculate the cross-domain sensing hysteresis time. potential gradient of resource load The degree of attenuation effect is determined, and the result is the cross-domain sensing hysteresis compensation factor. This cross-domain sensing hysteresis compensation factor The calculation method is as follows: ; in The preset hysteresis attenuation coefficient and A positive real number greater than 0, used to adjust the cross-domain sensing hysteresis time. The scaling factor for the degree of attenuation effect. The dimension is millisecond; hysteresis decay coefficient The typical value is 0.01 / ms. This value is set based on the fact that the cross-domain sensing hysteresis time... When it is on the order of 10ms The cross-domain sensing hysteresis compensation factor is approximately 0.1. Approximately 0.91, representing the resource load potential energy gradient. Approximately 91% remains effective, with a decay rate of about 9%; when the cross-domain sensing hysteresis time... When increased to the order of 100ms Approximately 1.0, Approximately 0.5, representing the potential energy gradient of resource load. Only about half of the data remains effective, with a significantly accelerated rate of degradation; when the cross-domain sensing lag time... Approaching 0 Approaching 1 indicates the potential energy gradient of resource load under ideal zero hysteresis conditions. No attenuation compensation is required;

[0046] The physical meaning of this calculation method lies in the cross-domain sensing hysteresis time. The larger the value, the longer the end-to-end time delay from collecting resource status information from the availability zone to obtaining that status from the scheduling decision center. The scheduling decision center calculates the resource load potential gradient based on this. The greater the deviation from the current true value, the more necessary it is to introduce a decay factor between 0 and 1 to quantify and compensate for this deviation; the cross-domain sensing hysteresis compensation factor This will serve as the basic intermediate variable for determining the hysteresis suppression coefficient in step 3.3. Its role is to transform the perceived hysteresis in the time dimension into a quantitative suppression degree on the scheduling convergence speed.

[0047] In a preferred embodiment of the present invention, step 3 above may further include:

[0048] Step 3.3: Based on the cross-domain sensing hysteresis compensation factor, determine the suppression ratio of cross-domain sensing hysteresis on scheduling convergence speed, and obtain the hysteresis suppression coefficient. Specifically, this includes: calculating the cross-domain sensing hysteresis compensation factor... Subsequently, based on this cross-domain sensing hysteresis compensation factor The hysteresis suppression coefficient is obtained by determining the proportion by which cross-domain sensing hysteresis inhibits scheduling convergence speed. The hysteresis suppression coefficient The calculation method is as follows ;in The range of values ​​is , The corresponding range of values ​​is The physical meaning of this calculation method lies in the cross-domain sensing hysteresis compensation factor. This represents the potential energy gradient of resource load under the condition of perception hysteresis. The proportion that remains effective, while the hysteresis inhibition coefficient This represents the potential energy gradient of the resource load. The proportion of information that is attenuated and suppressed due to sensory delay, the sum of which is always 1. .

[0049] Among them, the hysteresis inhibition coefficient The larger the value, the more significant the inhibitory effect of cross-domain perception lag on scheduling convergence speed. The scheduling decision engine should use a larger convergence weight when generating scheduling instructions to offset this inhibitory effect, in order to prevent the scheduling response speed from being excessively reduced due to perception lag. For example, when... hour This indicates that cross-domain perception lag has a 9% inhibitory effect on scheduling convergence speed, and the scheduling decision engine needs to increase the convergence weight accordingly to compensate for this inhibition; when hour This indicates that cross-domain perception lag leads to a potential gradient in resource load. Half of the information validity has been lost, and the scheduling decision engine needs to compensate with significantly larger convergence weights; this hysteresis suppression coefficient It will serve as the correction factor for the reverse weighted correction of resource load fluctuation amplitude in step 3.4. Its core function is to explicitly incorporate the adverse effects of the time lag dimension into the dynamic adjustment process of the convergence weight.

[0050] Step 3.4: Based on the hysteresis suppression coefficient and the resource load fluctuation amplitude, perform a reverse weighted correction on the resource load fluctuation amplitude to obtain the dynamic convergence weight. Specifically, this includes: calculating the hysteresis suppression coefficient... And the fluctuation range of resource load was calculated. Then, these two are used as input parameters for weighted correction, based on the hysteresis suppression coefficient. fluctuations in resource load Performing a reverse weighting correction yields the dynamically convergent weights. The dynamic convergence weight The calculation method is as follows: ; in The preset basic convergence weights and It is a positive real number between 0 and 1, with a typical value of 0.5. This value indicates that the scheduling decision engine performs cross-domain scheduling at a moderate convergence speed under conditions of imperceptible hysteresis and no load fluctuation, and can achieve an initial balance between scheduling aggressiveness and system stability. For resource load fluctuation range Normalized fluctuation amplitude after normalization and ; The maximum tolerable resource load fluctuation threshold is preset, with a typical value of 0.20. This value is set based on the fact that when the fluctuation of the overall resource utilization exceeds 20%, it indicates that the cross-domain scheduling operation has caused a relatively severe impact on the resource status of the availability zone. At this time, it is necessary to significantly increase the dynamic convergence weight to accelerate the rebalancing of the resource status.

[0051] To intuitively illustrate the adaptive adjustment logic of the dynamic convergence weights, see [link / reference]. Figure 4 The dynamic convergence weights shown in the figure are the response surface of hysteresis and fluctuation. This figure is displayed in a three-dimensional coordinate system. The X-axis is the cross-domain perceived hysteresis time, ranging from 0 to 2000ms, corresponding to the transmission delay span of cross-regional links; the Y-axis is the normalized resource load fluctuation amplitude, ranging from 0 to 1, corresponding to the intensity of historical scheduling oscillations; the Z-axis is the dynamic convergence weight, ranging from 0.5 to 0.95, corresponding to the convergence speed of scheduling decisions. Overall, the response surface shows a monotonically increasing trend: when both hysteresis time and fluctuation amplitude are low (X<200ms, Y<0.1), the weight is stable in the range of 0.5 to 0.55, corresponding to smooth scheduling in low uncertainty scenarios; when either hysteresis or fluctuation increases, the weight increases linearly; when the hysteresis reaches more than 1500ms and the fluctuation amplitude exceeds 0.8, the weight approaches the upper limit of 0.9, which is in line with the design logic of accelerating convergence in high uncertainty scenarios and can effectively avoid over-scheduling oscillations caused by traditional fixed weights from the root.

[0052] Further breaking down the surface features, the left half (X≤1000ms, i.e., scenarios within the same region or adjacent availability zones): the surface slope is relatively gentle, and for every 500ms increase in latency, the weight only increases by 0.05 to 0.08, indicating that latency has a limited impact on scheduling in short-link scenarios, and the weight is mainly dominated by resource fluctuation amplitude; the right half (X>1000ms, i.e., scenarios involving long-distance availability zones across cities): the surface slope becomes significantly steeper, and for every 500ms increase in latency, the weight increases by 0.12 to 0.18, indicating that latency becomes the core factor affecting scheduling stability in long-distance cross-domain scenarios, and the convergence weight needs to be reasonably increased to offset the decision bias caused by perception lag; the physical meaning of the calculation method is that the dynamic convergence weight... Based on the basic convergence weights Hysteresis-Volatile Joint Correction Term The product consists of two parts, representing the basic convergence weights. The baseline convergence rate under conditions of no perceived hysteresis and no load fluctuation was determined. The hysteresis-fluctuation joint correction term is then used to amplify and adjust the baseline convergence rate based on the degree of suppression of cross-domain perceived hysteresis and the severity of resource load fluctuation.

[0053] Specifically, when cross-domain sensing lag time When it is large, the hysteresis inhibition coefficient The corresponding increase makes the dynamic convergence weights Basic convergence weights This allows for a greater amplification factor, thereby compensating for scheduling response lag caused by perception delay with faster convergence speed; simultaneously, when resource load fluctuations... When the normalized fluctuation is large, The corresponding increase indicates that historical scheduling operations have had a significant impact on the resource status of the availability zone. In this case, it is also necessary to appropriately increase the dynamic convergence weight. This is to accelerate the scheduling convergence speed, so that the resource status of the availability zone can recover from the fluctuations to a balanced state as soon as possible.

[0054] For example, in , , In typical scenarios, dynamic convergence weights Compared to the basic convergence weights, the amplification is approximately 15%; in extreme scenarios, if and ,but The dynamic convergence weight is close to the upper limit of 1.0, indicating that the scheduling decision engine performs cross-domain scheduling at a near-maximum convergence speed to adapt to extreme perceived latency and load fluctuation conditions; ultimately, this dynamic convergence weight... This will serve as the core control parameter for the scheduling decision engine to generate cross-domain scheduling instructions in step 3.5. Its value directly determines the final cross-domain task scheduling amount determined in this round of scheduling decisions, i.e., the scheduling amount and the dynamic convergence weight. and resource load potential gradient The product of these factors is directly proportional, thus enabling adaptive and precise control of cross-domain scheduling intensity based on a comprehensive consideration of the impact of cross-domain perception lag and historical resource load fluctuation experience.

[0055] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Based on the resource load potential energy gradient, determine the potential energy driving direction and relative amplitude of the resource flow from the source availability zone to the target availability zone, and obtain the potential energy direction indication. Specifically, this includes: calculating the resource load potential energy gradient in step 2.5. Then, the scheduling decision engine determines the resource load potential gradient. Determine the potential energy driving direction and relative magnitude of resource flow from the source availability zone to the target availability zone; the result is the potential energy direction indicator. This potential energy direction indicator quantity The calculation method is as follows That is, the potential energy direction indicator. The value is directly taken from the output resource load potential energy gradient. The direction of potential energy driving is indicated by the potential energy direction indicator. The sign bit is determined when A positive value indicates that resources should migrate from the source availability zone to the target availability zone. A negative value indicates that resources should migrate in the reverse direction from the target availability zone to the source availability zone. absolute value Less than the preset potential energy direction dead zone threshold and When the typical value is 0.01, it means that the resource load potential difference between the source availability zone and the target availability zone has basically disappeared. At this time, the scheduling decision engine determines that the system is in a balanced state and there is no need to perform this round of cross-domain scheduling operation.

[0056] The relative amplitude driven by potential energy is indicated by the direction of potential energy. absolute value Sure, The larger the value, the more significant the difference in resource load potential between the source and target availability zones, and the larger the scale of resources that need to be scheduled should be. The smaller the value, the more balanced the resource load of the two availability zones is; in determining the potential energy direction indicator... Then, it is compared with the historical potential energy direction indication corresponding to the previous scheduling cycle. Perform a comparison, if and If the sign bits are the same, it indicates that the potential energy direction of the cross-domain resource load imbalance has not reversed, and the current round of scheduling can continue in the same direction. If the sign bits are opposite, it indicates that the potential energy direction has reversed, and the scheduling decision engine will trigger a direction reversal suppression mechanism to adjust the potential energy direction indicator of the current round. The amplitude multiplied by a direction reversal attenuation coefficient less than 1 Reduce the magnitude. The typical value is 0.5 to avoid frequent reversals of the scheduling direction from impacting system stability; this potential energy direction indicator... It will serve as a directional reference input for calculating the target migration amount in subsequent steps, and its role is to provide a clear scheduling direction and magnitude basis for the subsequent migration amount conversion.

[0057] Step 4.2: Based on the potential energy direction indicator and the dynamic convergence weight, apply convergence constraints to the potential energy direction indicator and convert it into the number of tasks to be migrated to obtain the target migration amount. Specifically, this includes: calculating the potential energy direction indicator... And the dynamic convergence weights are calculated. Then, using these two as core input parameters, based on the dynamic convergence weights... Indicator of potential energy direction A convergence constraint is applied to the magnitude of the constraint, and the result after constraint is converted into the number of tasks to be transferred. The result is the target transfer amount. The target migration amount The calculation method is as follows ;in The preset baseline number of migration tasks, symbol This indicates a rounding down operation to ensure the target migration count is an integer number of tasks; the baseline migration task count. The value is determined by the scheduling decision engine based on the total number of migrated task instances in the compute resource pools of the source and target availability zones. preset ratio To determine, that is ; among which the preset ratio The value range is [3%, 10%], with a typical value of 5%; during the system initialization phase. Take the arithmetic mean of the total number of migrated task instances in both the source and target availability zones; if the calculated... Less than the preset baseline migration task limit ( (The typical value is 10 tasks), then... Set as This ensures that the baseline number of migration tasks still has practical scheduling significance when the total number of migrateable tasks is small.

[0058] When the dynamic convergence weight The absolute value of the potential energy direction indicator product Less than the preset minimum migration ratio threshold and When the typical value is 0.01, the target migration amount is... Forcing a value of 0 indicates that the current cross-domain resource load imbalance is insufficient to trigger a substantial task migration operation, thus avoiding the problem of disproportionate system overhead and scheduling benefits caused by performing a small amount of scheduling; when the calculated value is... Greater than the preset maximum migration amount per round and When the typical value is 500 tasks, the scheduling decision engine will determine the target migration amount. Cut off to This is to prevent a single round of scheduling operations from migrating too many tasks and causing a sudden resource impact on the target availability zone; the physical meaning of the calculation method is that the dynamic convergence weight... The scheduling convergence speed is adaptively adjusted from two dimensions: time hysteresis and load fluctuation. Potential energy direction indicator. absolute value The number of baseline migration tasks reflects the degree of imbalance in cross-domain resource load from a spatial perspective. The system capacity dimension provides a benchmark for task conversion. The target migration amount is obtained by multiplying the three dimensions and rounding down and applying upper and lower bound constraints. It can comprehensively consider multiple influencing factors and accurately calculate the number of tasks to be migrated for each cross-domain scheduling operation.

[0059] Step 4.3: Based on the target migration amount, generate decision instructions to drive cross-availability zone resource scheduling, obtaining adaptive balancing scheduling instructions. Specifically, this includes: calculating the target migration amount... Furthermore, after step 4.1 clarifies the potential energy driving direction, the scheduling decision engine determines the target migration amount. Generate decision instructions to drive cross-availability zone resource scheduling; that is, first based on the potential energy direction indication. The sign bit determines the source availability zone identifier and target availability zone identifier for this round of scheduling. When a positive value is taken, the resource supply status determined in step 2.4 will be used. The corresponding availability zone is identified as the source availability zone, and the resource demand status determined in step 2.4 is then processed. The corresponding availability zone is determined as the target availability zone, when When a negative value is taken, the above correspondence is reversed, that is, the availability zone with the original resource supply status is changed to the target availability zone, and the availability zone with the original resource demand status is changed to the source availability zone.

[0060] Further, according to the global identifier naming rules for availability zones described in step 1.1, the source availability zone network address corresponding to the source availability zone identifier and the target availability zone network address corresponding to the target availability zone identifier are queried from the availability zone registration information table. The source availability zone identifier, target availability zone identifier, source availability zone network address, target availability zone network address, and target migration amount are then combined. The instruction is encapsulated as a structured adaptive balancing scheduling instruction. This instruction includes an instruction generation timestamp, a unique identifier, a validity period, and an integrity checksum. The timestamp is the system clock reading when the scheduling decision engine generates the instruction. The unique identifier is composed of the scheduling decision engine identifier, the scheduling round number, and a random number. The typical validity period is 30 seconds to prevent expired instructions from being executed repeatedly. The integrity checksum is generated by performing a digest calculation on all preceding fields of the instruction using a cyclic redundancy check algorithm. The adaptive balancing scheduling instruction is written to the instruction distribution queue and pushed asynchronously to the task scheduler in the source availability zone via a cross-availability zone communication link. Simultaneously, an instruction confirmation timeout timer is started. If no confirmation is received from the source availability zone within the validity period, a retransmission mechanism is triggered. The generation of the adaptive balancing scheduling instruction marks the transition of this round of cross-domain scheduling decision-making from the analysis and calculation phase to the instruction execution phase, carrying the target migration amount. This directly determines the actual execution level of this round of scheduling operations.

[0061] In a preferred embodiment of the present invention, step 4 above may further include:

[0062] Step 4.4: Based on the adaptive balancing scheduling instruction, the source availability zone identifier and target availability zone identifier carried in the adaptive balancing scheduling instruction are parsed and extracted to obtain scheduling routing information. Specifically, this includes: after receiving the adaptive balancing scheduling instruction and pushing it through the cross-availability zone communication link, the task scheduler executor deployed locally in the source availability zone extracts the adaptive balancing scheduling instruction from the instruction receiving queue and performs parsing processing; firstly, the instruction is subjected to integrity verification, specifically by recalculating the cyclic redundancy check value of all fields in the instruction except the check code field, comparing the recalculated check value with the instruction integrity check code carried in the instruction bit by bit. If the two are consistent, it is determined that there is no data corruption during the instruction transmission process, and the verification passes; if the two are inconsistent, it is determined that the instruction has been tampered with or lost during transmission, and the task scheduler executor discards the instruction and returns a negative confirmation message of instruction verification failure to the scheduling decision engine.

[0063] After successful verification, the task scheduler further checks whether the difference between the instruction generation timestamp and the current system clock exceeds the instruction's validity period. If it does, the instruction is deemed expired and discarded, and a negative confirmation message of instruction expiration is returned to the scheduling decision engine. After passing the timeliness check, the task scheduler extracts the source availability zone identifier, target availability zone identifier, source availability zone network address, target availability zone network address, and target migration amount from the instruction's structured fields. The extracted source availability zone identifier and target availability zone identifier are combined to form scheduling routing information. This scheduling routing information explicitly indicates the starting and destination endpoints of this round of cross-domain scheduling operations. The source availability zone identifier is used to locate the residence location of the task to be migrated within the local availability zone, and the target availability zone identifier and its corresponding target availability zone network address are used to determine the destination network addressing path of the task to be migrated during cross-domain transmission. This scheduling routing information is then combined with the target migration amount... The data is then passed to the local task filtering module and a confirmation receipt is returned to the scheduling decision engine, providing precise routing and quantity constraints for the selection of the set of tasks to be migrated in step 4.5.

[0064] Step 4.5: Based on the scheduling routing information and the target migration amount, filter stateless computing tasks or risk control verification requests that match the target migration amount in quantity from the source availability zones indicated by the source availability zone identifier, to obtain a set of tasks to be migrated. Specifically, this includes: parsing the obtained scheduling routing information and target migration amount... Subsequently, the task scheduler executor deployed locally in the source availability zone selects the target migration volume from the computing resource pool managed by that source availability zone based on the source availability zone identifier in the scheduling routing information. A set of tasks to be migrated is obtained by matching the number of stateless computation tasks or risk control verification requests. First, all task instances in the local task running status table of the source availability zone that are in a running or waiting state are traversed. Computation tasks marked as stateless and risk control verification requests marked as cross-domain are filtered out based on task type tags. All selected candidate task instances are aggregated into a candidate migration task pool, and the total number of task instances in this pool is counted. .

[0065] like If the value is 0, it indicates that there are currently no migrateable task instances in the source availability zone. The task scheduler executor directly returns a negative acknowledgment message indicating that there are no migrateable tasks to the scheduling decision engine, and this round of scheduling operations terminates. If the value is greater than 0, then the total number of task instances in the candidate migration task pool will be [number]. With target migration amount Compare the two values ​​and take the smaller value as the actual number of migrations in this round. ,Right now when Greater than or equal to At that time, select the top-performing migration task from the candidate migration task pool in order of task submission time. The task set consists of several task instances. If multiple task instances have the same submission time, they are further sorted lexicographically by task instance identifier to ensure the determinism of the selection result; when Less than At that time, all candidates in the migration task pool will be selected. Each task instance is included in the set of tasks to be migrated. At this time, the actual number of tasks to be migrated in this round of scheduling operation is capped at the number of stateless tasks actually available in the source availability zone, so as to avoid the failure of some scheduling instructions due to the target migration amount exceeding the actual migrateable task capacity. After determining the set of tasks to be migrated, the task instance identifier of each task instance in the set is recorded in the scheduling operation log of this round, so that the scheduling decision engine can update and trace the historical scheduling instruction sequence in subsequent scheduling cycles.

[0066] Step 4.6: Based on the set of tasks to be migrated and the scheduling routing information, schedule each task instance in the set of tasks to be migrated from the source availability zone indicated by the source availability zone identifier to the target availability zone indicated by the target availability zone identifier, thereby completing cross-availability zone adaptive load balancing. Specifically, after filtering and obtaining the set of tasks to be migrated and parsing the scheduling routing information, the task scheduling executor deployed locally in the source availability zone schedules each task instance in the set of tasks to be migrated from the source availability zone indicated by the source availability zone identifier to the target availability zone indicated by the target availability zone identifier, based on the set of tasks to be migrated and the scheduling routing information. Specifically, first, perform a state freeze operation on each task instance in the set of tasks to be migrated, marking the running status of the task instance from running or waiting to migration suspended state, so as to suspend the continued execution of the task instance in the source availability zone. At the same time, perform a consistent snapshot of the memory data pages, register snapshots, and consumption offsets of processed input data currently held by the task instance, and merge and store the snapshot results with the task context data of the task instance.

[0067] The task context data, intermediate execution states, task dependency graphs, and aforementioned consistency snapshots of each task instance in the task set to be migrated are further serialized into migration data packets. The serialization format adopts a cross-availability zone data exchange protocol agreed upon with the target availability zone. The migration data packet also includes a data packet sequence number, total data packet length, and data packet integrity check code. The migration data packet is transmitted packet by packet from the source availability zone to the target availability zone pointed to by the network address of the target availability zone in the scheduling routing information through a cross-availability zone communication link. A reliable transmission protocol with retransmission confirmation is used during transmission to ensure that the migration data packet is not lost under unreliable network conditions. After the target availability zone successfully receives all migration data packets and verifies that the data packet integrity check code passes one by one, the task recovery executor in the target availability zone deserializes and restores the migration data packets. Based on the task context data and consistency snapshot, the corresponding task instance is recreated in the target availability zone, and the running status of the task instance is restored from the migration suspended state to the running or waiting state, so that the task instance can continue to execute the subsequent calculation logic in the target availability zone from the execution breakpoint recorded in the snapshot.

[0068] Once all task instances in the set to be migrated have successfully resumed execution in the target availability zone, the task recovery executor in the target availability zone sends a migration completion confirmation message to the task scheduling executor in the source availability zone. Upon receiving the confirmation message, the task scheduling executor releases and reclaims the local resources occupied by the migrated task instances in the source availability zone and reports the confirmation message of the completion of this round of cross-domain scheduling execution to the scheduling decision engine, thus completing a complete cross-availability zone adaptive load balancing closed-loop operation. If an abnormal situation such as network interruption or target availability zone recovery execution failure occurs during the migration data packet transmission, a migration rollback mechanism will be triggered to restore the frozen but not yet successfully migrated to the target availability zone to the running state before migration in the source availability zone, ensuring that the atomicity of the cross-domain scheduling operation and the consistency of task execution are not compromised.

[0069] like Figure 2 As shown, embodiments of the present invention also provide an adaptive balanced scheduling closed-loop optimization system for cross-regional multi-availability zone operation and maintenance architectures, including: The data acquisition and hysteresis perception module is used to collect multi-dimensional resource status data and collection timestamps of each availability zone in real time, persist historical multi-dimensional resource status data, fit the natural evolution trend of local services based on the historical multi-dimensional resource status data, and calculate the difference in collection decision time and the pre-stored cross-domain link round-trip latency to obtain the cross-domain perception hysteresis time. The state compensation and gradient calculation module is used to backtrack the resource state snapshot at the actual historical acquisition time from the historical multidimensional resource state data based on the cross-domain sensing lag time, and to perform time alignment compensation on the resource state snapshot in combination with the natural evolution trend to obtain the compensated multidimensional resource state data, and to calculate the resource supply and demand difference between the source availability zone and the target availability zone to obtain the resource load potential gradient. The fluctuation assessment and weight generation module is used to calculate the resource load fluctuation amplitude based on the pre-stored historical scheduling instruction sequence, calculate the cross-domain sensing hysteresis compensation factor based on the resource load potential energy gradient and the cross-domain sensing hysteresis time, and generate dynamic convergence weights based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude. The migration decision and scheduling execution module is used to calculate the target migration amount based on the resource load potential gradient and the dynamic convergence weight, generate an adaptive balance scheduling instruction, and dynamically schedule the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone to the target availability zone based on the adaptive balance scheduling instruction and the target migration amount.

[0070] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0071] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0072] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0073] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture, characterized in that, The method includes: Real-time collection of multi-dimensional resource status data and collection timestamps from each availability zone is used to persist historical multi-dimensional resource status data. Based on this historical data, the natural evolution trend of local services is fitted, following the following fitting calculation formula: [The formula is missing from the original text, so a direct translation isn't possible.] Within a time window consisting of several collection cycles, let the first... Each collection timestamp is The corresponding local business load metric observation value is ,in The value ranges from 1 to integers and The time window length parameter is pre-configured; then the fitted natural evolution trend function is obtained. satisfy ,in Let be the trend slope parameter, and its expression is as follows: ； The baseline intercept parameter and The above summation symbols All expressed their opinion From 1 to Perform iterative summation to obtain the trend slope parameter. molecular part The product of window length and total time-load covariance is subtracted from the product of total time and total load. This difference, after normalizing the denominator, yields the optimal estimate of the linear correlation coefficient between time and load in the least squares sense. The denominator of the linear regression slope calculation reflects the dispersion of the time independent variable within the window; The positive and negative signs indicate the long-term growth trend and long-term decay trend of the load relative to the time axis, respectively. The absolute value of the baseline intercept parameter indicates the rate of change in the trend. molecular part This represents the cumulative total load minus the product of the trend slope parameter and the cumulative total over time, with the difference calculated over a window of length. After normalization, the baseline load level, i.e., the intercept value, corresponding to the start time of the current sliding time window is obtained. The value reflects the baseline load level, i.e., the intercept value, corresponding to the start time of the current sliding time window; the natural evolution trend function This is an evolution curve that conforms to historical change patterns and can be extrapolated along the time axis, used to describe the natural change trend of local business load over time under the condition of no external scheduling intervention; the difference between the current scheduling decision time and the collection timestamp is calculated synchronously to obtain the collection decision time difference value; the collection decision time difference value is added to the pre-stored cross-domain link round-trip delay to obtain the cross-domain perception lag time. Based on the cross-domain sensing lag time, a snapshot of the resource status at the actual historical acquisition time is obtained by backtracking from historical multidimensional resource status data; time alignment compensation is performed based on the resource status snapshot and the natural evolution trend to obtain compensated multidimensional resource status data; based on the compensated multidimensional resource status data, the resource supply and demand difference between the source availability zone and the target availability zone is calculated to obtain the resource load potential gradient. Based on the pre-stored historical scheduling instruction sequence, the resource load fluctuation amplitude is statistically obtained; based on the resource load potential energy gradient and cross-domain sensing hysteresis time, the cross-domain sensing hysteresis compensation factor is calculated; based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude, the dynamic convergence weight is obtained. Based on the resource load potential gradient and dynamic convergence weight, the target migration amount and adaptive balancing scheduling instructions are calculated. According to the adaptive balancing scheduling instructions and the target migration amount, the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone are dynamically scheduled to the target availability zone.

2. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 1, characterized in that, Real-time collection of multi-dimensional resource status data and collection timestamps from each availability zone, and persistence of historical multi-dimensional resource status data; fitting of historical multi-dimensional resource status data to obtain the natural evolution trend of local services; The difference between the current scheduling decision time and the collection timestamp is calculated synchronously to obtain the collection decision time difference value; The difference in acquisition and decision-making time is added to the pre-stored cross-domain link round-trip delay to obtain the cross-domain sensing hysteresis time, including: Based on the collection probes deployed in each availability zone, multidimensional resource status data is acquired in real time, and each piece of multidimensional resource status data is marked with a corresponding collection timestamp to form multidimensional resource status data marked with collection timestamps. The multidimensional resource status data marked with the collection timestamp is persistently stored in time series to obtain historical multidimensional resource status data; Based on historical multidimensional resource status data, the trend characteristics of local business load changes over time are extracted, and the natural evolution trend of local business is obtained by fitting. Calculate the time difference between the current scheduling decision time and the collection timestamp to obtain the collection decision time difference value; Based on the pre-stored network link quality information, the round-trip time between the source availability zone and the target availability zone is extracted to obtain the pre-stored cross-domain link round-trip time. The difference in the acquisition and decision-making time is added to the pre-stored cross-domain link round-trip delay to obtain the cross-domain sensing lag time.

3. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 2, characterized in that, Based on the cross-domain sensing lag time, a snapshot of the resource status at the actual historical acquisition time is obtained by backtracking from historical multidimensional resource status data; time alignment compensation is performed based on the resource status snapshot and the natural evolution trend to obtain compensated multidimensional resource status data; Based on the compensated multidimensional resource status data, the resource supply and demand difference between the source availability zone and the target availability zone is calculated to obtain the resource load potential gradient, including: Based on the cross-domain sensing hysteresis time, determine the historical backtracking time offset; Based on the historical backtracking time offset and the historical multidimensional resource status data, the multidimensional resource status at the corresponding moment is located and extracted to obtain a snapshot of the resource status at the actual historical acquisition moment. Based on the resource status snapshot and the natural evolution trend, forward compensation is performed on the resource status snapshot along the time axis to obtain compensated multidimensional resource status data. Based on the compensated multidimensional resource status data, the resource supply status of the source availability zone and the resource demand status of the target availability zone are determined respectively. Based on the resource supply status of the source availability zone and the resource demand status of the target availability zone, the resource supply and demand difference between the resource supply status and the resource demand status is calculated to obtain the resource load potential gradient.

4. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 3, characterized in that, Based on the pre-stored historical scheduling instruction sequence, the resource load fluctuation amplitude is statistically obtained; based on the resource load potential gradient and cross-domain sensing hysteresis time, the cross-domain sensing hysteresis compensation factor is calculated, including: Based on the pre-stored historical scheduling instruction sequence, extract the scheduling action type and corresponding scheduling quantity from the historical scheduling instruction sequence at each historical scheduling moment, and statistically analyze the range of resource load changes in the source availability zone and the target availability zone before and after scheduling to obtain the resource load fluctuation amplitude. Based on the resource load potential energy gradient and the cross-domain sensing hysteresis time, the degree of attenuation of the cross-domain sensing hysteresis time on the resource load potential energy gradient is calculated, and the cross-domain sensing hysteresis compensation factor is obtained.

5. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 4, characterized in that, Based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude, dynamic convergence weights are obtained, including: Based on the cross-domain sensing hysteresis compensation factor, the suppression ratio of cross-domain sensing hysteresis on scheduling convergence speed is determined, and the hysteresis suppression coefficient is obtained. Based on the hysteresis suppression coefficient and the resource load fluctuation amplitude, the resource load fluctuation amplitude is reverse-weighted and corrected to obtain the dynamic convergence weight.

6. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 5, characterized in that, Based on the resource load potential gradient and dynamic convergence weights, the target migration amount and adaptive balancing scheduling instructions are calculated, including: Based on the resource load potential energy gradient, the potential energy driving direction and relative amplitude of the resource flow from the source availability zone to the target availability zone are determined, and the potential energy direction indication is obtained. Based on the potential energy direction indicator and the dynamic convergence weight, the potential energy direction indicator is applied a convergence constraint and then converted into the number of tasks to be migrated, thus obtaining the target migration amount. Based on the target migration amount, decision instructions are generated to drive cross-availability zone resource scheduling, resulting in adaptive balancing scheduling instructions.

7. The adaptive balanced scheduling closed-loop optimization method for cross-regional multi-availability zone operation and maintenance architecture according to claim 6, characterized in that, Based on the adaptive balancing scheduling instructions and the target migration amount, the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone are dynamically scheduled to the target availability zone, including: Based on the adaptive balancing scheduling instruction, the source availability zone identifier and target availability zone identifier carried in the adaptive balancing scheduling instruction are parsed and extracted to obtain the scheduling routing information; Based on the scheduling routing information and the target migration amount, stateless computing tasks or risk control verification requests that match the target migration amount in quantity are filtered from the source availability zones indicated by the source availability zone identifier to obtain a set of tasks to be migrated. Based on the set of tasks to be migrated and the scheduling routing information, each task instance in the set of tasks to be migrated is scheduled from the source availability zone indicated by the source availability zone identifier to the target availability zone indicated by the target availability zone identifier, thereby completing cross-availability zone adaptive load balancing.

8. An adaptive balanced scheduling closed-loop optimization system for cross-regional multi-availability zone operation and maintenance architecture, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and hysteresis perception module is used to collect multi-dimensional resource status data and collection timestamps of each availability zone in real time, persist historical multi-dimensional resource status data, fit the natural evolution trend of local services based on the historical multi-dimensional resource status data, and calculate the difference in collection decision time and the pre-stored cross-domain link round-trip latency to obtain the cross-domain perception hysteresis time. The state compensation and gradient calculation module is used to backtrack the resource state snapshot at the actual historical acquisition time from the historical multidimensional resource state data based on the cross-domain sensing lag time, and to perform time alignment compensation on the resource state snapshot in combination with the natural evolution trend to obtain the compensated multidimensional resource state data, and to calculate the resource supply and demand difference between the source availability zone and the target availability zone to obtain the resource load potential gradient. The fluctuation assessment and weight generation module is used to calculate the resource load fluctuation amplitude based on the pre-stored historical scheduling instruction sequence, calculate the cross-domain sensing hysteresis compensation factor based on the resource load potential energy gradient and the cross-domain sensing hysteresis time, and generate dynamic convergence weights based on the cross-domain sensing hysteresis compensation factor and the resource load fluctuation amplitude. The migration decision and scheduling execution module is used to calculate the target migration amount based on the resource load potential gradient and the dynamic convergence weight, generate an adaptive balance scheduling instruction, and dynamically schedule the corresponding number of stateless computing tasks or risk control verification requests in the source availability zone to the target availability zone based on the adaptive balance scheduling instruction and the target migration amount.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images