A control method and system for a rack-mounted cluster server

Abstract

The application discloses a kind of control method and system for mounting type cluster server, it is related to computer technical field, the method is obtained by calling power monitoring interface supply voltage original sequence, and executes time reference alignment processing to form supply voltage sequence set;Screening effective operation section, extract power supply voltage segment to carry out change direction identification and continuous section detection, construct power supply change section sequence, and generate power supply rhythm feature sequence;Identify abnormal candidate segment and construct power supply drop pre-judgment window;When abnormal candidate segment enters pre-judgment window, call task metadata and construct task classification table, respectively execute parallel convergence, task splitting, staggered scheduling and freezing control for different types of tasks;After power supply recovers to stable form, the execution state of task is sorted and recovered, time stamp alignment and task segment merging.The application constructs power supply drop pre-judgment and scheduling collaborative mechanism by power supply rhythm identification and task scheduling strategy linkage.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a control method and system for mounted cluster servers. Background Technology

[0002] Servers, as the core computing carriers of information systems, are typically deployed in data centers or server rooms, relying on stable power and constant network conditions for continuous operation. With the increasing demand for edge computing and on-site deployment, servers have gradually evolved from centralized server room setups to mounted cluster servers that can be installed in vehicles, industrial equipment, or mobile platforms. These servers operate in mobile or semi-fixed scenarios, with computing nodes working collaboratively in a cluster. However, their power supply often depends on on-site power supplies, inverters, or temporary power interfaces, exhibiting fluctuating and discontinuous power supply characteristics. Since mounted cluster servers need to maintain high-density computing power while operating in non-standard power environments, coordinating the control of their power supply rhythm characteristics and task scheduling behavior—developing control methods for mounted cluster servers—has become a key technical challenge in mobile computing scenarios.

[0003] Existing server-oriented control strategies typically assume relatively stable power conditions. Scheduling systems primarily allocate resources based on CPU, memory, and network bandwidth, without structurally analyzing the rhythmic changes in power supply voltage. In a cluster server environment, when voltage experiences short-cycle dips, recoveries, or oscillations, the scheduling system still executes tasks according to the predetermined concurrency strategy, lacking a forward-looking discrimination and categorized intervention mechanism based on power supply rhythm. Furthermore, existing solutions fail to model the correlation between power supply variation periods and task resource usage patterns, and do not establish progressive control logic between the dip prediction window and task scheduling, resulting in a disconnect between task scheduling and power supply status. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a control method and system for mounted cluster servers, solving the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a control method for mounted cluster servers, comprising the following steps: S1. Call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment, form a power supply voltage sequence set and filter the valid operating segments to extract the power supply voltage segments. S2. By constructing a power supply change segment sequence, numbering each change segment in sequence and calculating the duration of adjacent segments, a power supply rhythm characteristic sequence is generated, and historical stable operation power supply segments are retrieved to obtain a set of historical stable rhythms. S3. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set to construct a power supply drop prediction window. Generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. S4. When an abnormal candidate segment enters the fall prediction window, obtain the task queue status and execute the beat scheduling process according to the task type. After power is restored, execute task sorting, status verification and task segment merging.

[0006] Preferably, S1 includes S11; S11. Call the power monitoring interface at the operating system layer of the cluster server computing node to read the node power supply voltage sampling data, form the original power supply voltage sequence, add a timestamp to the original power supply voltage sequence of each node, and then perform time base alignment processing based on the unified time source of the cluster to unify the original power supply voltage sequences of all nodes to the same time coordinate system, and output the power supply voltage sequence set.

[0007] Preferably, S1 further includes S12; S12. Perform sequence segmentation on the power supply voltage sequence set according to the continuous time window rule, construct several continuous time segments, and perform running status filtering in each time segment, specifically: whether the number of sampling points is lower than the set minimum sampling number, whether the timestamp is in reverse order, and whether the node status log is in restart, shutdown or maintenance state. If any condition is met within each consecutive time segment, the current time segment is marked as an invalid segment. If no condition exists within each consecutive time segment, the current time segment is marked as a valid operating segment, and the time segment that meets the valid operating segment is determined as the power supply voltage segment.

[0008] Preferably, S2 includes S21; S21. The voltage sampling sequence in the power supply voltage segment is traversed according to the sampling time order. The voltage values ​​of the current sampling point and the previous sampling point are read. The change direction of the current sampling point is determined according to the voltage magnitude rule between consecutive sampling points. Specifically, when the voltage of the current sampling point is higher than that of the previous sampling point, it is marked as an upward change. When the voltage of the current sampling point is lower than that of the previous sampling point, it is marked as a downward change. When the difference between the two is within the sampling allowable error range, it is marked as a stable change. A continuous segment detection process is performed on the change direction identifier sequence. When the change direction identifiers of adjacent sampling points are consistent, the sampling intervals in the current consistent state are registered as the same change operation segment, and the start and end sampling positions of the current operation segment are recorded. When a change in the change direction identifier of adjacent sampling points is detected, the sampling position where the direction changes is determined as the change boundary position, and multiple change segments within the power supply voltage segment are combined to form a power supply change segment sequence.

[0009] Preferably, S2 further includes S22; S22. Number each change segment in the power supply change segment sequence according to the order of appearance, and record the duration interval between adjacent segments to generate a power supply rhythm feature sequence. At the same time, retrieve the stable operation power supply segments saved during the historical operation of the node, and generate a historical rhythm feature sequence using the same processing method to obtain a historical stable rhythm set.

[0010] Preferably, S3 includes S31; S31. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set, as follows; When the arrangement of the changing segments in the power supply rhythm characteristic sequence is consistent with the historical stable rhythm set, the corresponding power supply voltage segment is marked as a stable power supply segment. The voltage data in each change segment of the power supply rhythm characteristic sequence is compared point by point. When the voltage of multiple sampling points shows a monotonically decreasing trend and the decrease is greater than the set decrease threshold, the current change segment is marked as a decreasing change segment. When the voltage data after the decreasing change segment reverses and rises, and the rise is greater than the set rise threshold, the current change segment is marked as a recovery change segment. When a combination of decreasing and recovering change segments is detected in the current power supply rhythm characteristic sequence, and the current combination does not appear in the historical stable rhythm set, the corresponding power supply voltage segment is identified as an abnormal candidate segment. The starting point of the abnormal candidate segment combination is used as the window start point, and the position where the change recovery is completed is used as the window end point to generate a power supply drop prediction window. The threshold for the decrease is set based on the maximum decrease in normal power supply fluctuations in historical data, and the threshold for the recovery is set based on the minimum effective recovery amplitude of the voltage recovery process in historical data.

[0011] Preferably, S3 further includes S32; S32. Obtain metadata of tasks to be executed and tasks currently being executed within the onboard cluster server through the Kubernetes scheduler. The metadata includes task type, resource consumption, execution time, CPU usage, and I / O frequency. Use a classification algorithm to classify the resource usage patterns of the metadata into continuous occupation type, burst type, synchronous sensitive type, and segmentable type, and generate a task classification table containing a unique identifier, task type, required resources, and lifecycle data for each task.

[0012] Preferably, S4 includes S41; S41. When an abnormal candidate segment of a node enters the fall prediction window, extract the task queue status from the current node's task classification table and start the beat scheduling process, as follows; To identify burst-type tasks, a parallelism convergence algorithm is used to merge multiple parallel tasks, and a batch processing algorithm is used to merge several tasks into a single task. Identify segmentable tasks and preset task segmentation thresholds based on the shortest execution task during historical stable operation. When each execution segment in a segmentable task exceeds the task segmentation threshold, a task splitting algorithm is used to split it into several sub-execution segments. The length of each execution segment is dynamically adjusted according to the power supply fluctuation cycle. Identify synchronization-sensitive tasks and use a dynamic synchronization-triggered scheduling algorithm to stagger cross-node tasks. Identify persistently occupied tasks, freeze the current task using a protective freeze algorithm, and resume when voltage fluctuations return to stability.

[0013] Preferably, S4 further includes S42; S42. After the power supply is restored to a stable state, tasks are sorted and scheduled based on the historical resource scheduling analysis algorithm. When a task that needs to be synchronized across nodes is detected, the synchronization alignment point of the task is restored through the timestamp synchronization mechanism. The task segments are checked one by one according to the log analysis algorithm to confirm the execution status of each task segment. The task segments are then merged according to the execution order and timestamp of the tasks.

[0014] A control system for mounted cluster servers includes a power supply rhythm acquisition module, an anomaly identification module, a drop prediction module, and a fingerprint database write-back module. The power supply rhythm acquisition module is used to call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment processing, a power supply voltage sequence set is formed, and the effective operating segment is selected to extract the power supply voltage segment. The anomaly identification module is used to construct a power supply change segment sequence, number each change segment sequentially and calculate the duration of adjacent segments, generate a power supply rhythm feature sequence, and retrieve historical stable operation power supply segments to obtain a set of historical stable rhythms. The power supply drop prediction module is used to perform sequence alignment and comparison processing on the power supply rhythm feature sequence and the historical stable rhythm set, construct the power supply drop prediction window, and generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. The fingerprint database write-back module is used to obtain the task queue status and execute the clock scheduling process according to the task type when abnormal candidate segments enter the fall prediction window. After power is restored, it performs task sorting, status verification and task segment merging.

[0015] This invention provides a control method and system for mounted cluster servers. It has the following beneficial effects: (1) This method adds timestamps to the original power supply voltage sequence of each node and performs time reference alignment processing based on the unified time source of the cluster. All node data are unified into the same time coordinate system, eliminating the interference of cross-node sampling deviation on rhythm analysis. By determining the minimum number of samples, detecting the reverse order of timestamps, and verifying the node status log, the effective operating segments are screened, so that the power supply voltage segments only come from the real operating state interval. The power supply voltage segments are marked with change direction, registered with change operating segments, numbered with change segments, and recorded with duration intervals to construct the power supply rhythm feature sequence. The historical rhythm feature sequence is generated in the same way to form a set of historical stable rhythms. The combination of decreasing change segments and recovering change segments is identified by sequence alignment and comparison processing, and a power supply drop prediction window is generated based on the set decrease amplitude threshold and recovery amplitude threshold. This link completes the entire process organization from original voltage sampling to the determination of abnormal candidate segments and window boundaries, so that the power supply fluctuation is transformed from discrete voltage values ​​into rhythmic expressions with structural order significance, and realizes the windowed definition of abnormal combination structures.

[0016] (2) After the power supply drop prediction window is triggered, this method constructs a structural response path on the task side. Task metadata, including task type, resource consumption, execution time, CPU usage, and I / O frequency, is obtained through the Kubernetes scheduler. This metadata is then categorized according to resource usage patterns into continuous occupancy, burst type, synchronous sensitive type, and segmentable type, generating a task classification table containing unique identifiers, task types, required resources, and lifecycle data. When an abnormal candidate segment enters the power supply drop prediction window, the task queue status in the task classification table is extracted, and the tick scheduling process is initiated: for burst type tasks, parallel convergence and batch processing algorithms are executed; for segmentable tasks, the shortest execution task in historical stable operation is pre-defined to a task segmentation threshold, and the execution segment length is adjusted according to the power supply fluctuation cycle; for synchronous sensitive tasks, a dynamic synchronous trigger scheduling algorithm is executed to stagger peak times; and for continuous occupancy tasks, a protective freeze algorithm is executed. The above processing is based on the power supply rhythm characteristic sequence and the power supply drop prediction window, enabling the task scheduling logic to correspond with the power supply change segment structure, thus completing the categorized handling of the task queue status during the power supply anomaly phase.

[0017] (3) After the power supply is restored to a stable power supply segment, the method performs a convergence processing on the scheduling behavior within the window period. The task is sorted by the historical resource scheduling analysis algorithm. When there are cross-node synchronization tasks, the synchronization alignment point is restored by the timestamp synchronization mechanism. The task execution status is checked segment by segment according to the log analysis algorithm. The task segments are merged according to the task execution order and timestamp. At the same time, at the structural level, a complete closed loop is formed by the power supply rhythm acquisition module, the anomaly identification module, the drop prediction module and the fingerprint database write-back module, so that the data transmission between the power supply voltage sequence set, the power supply rhythm feature sequence, the task classification table and the task queue status is consistent. Compared with the scheduling method that only relies on the instantaneous voltage threshold trigger, this scheme completes the entire process organization of power supply rhythm identification, anomaly candidate segment judgment, power supply drop prediction window construction, task type scheduling and task segment merging in the recovery stage, so that the mounted cluster server has the continuous control capability from identification to recovery in the power supply fluctuation scenario. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a control method for a mounted cluster server according to the present invention. Figure 2 This is a schematic diagram of the control system for a mounted cluster server according to the present invention; Figure 3 This is a block diagram illustrating the logic principle of a control method for a mounted cluster server according to the present invention. Detailed Implementation

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

[0020] Example 1 Please see Figure 1 This invention provides a control method for mounted cluster servers. To achieve the above objectives, this invention employs the following technical solution, comprising the following steps: S1. Call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment, form a power supply voltage sequence set and filter the valid operating segments to extract the power supply voltage segments. S2. By constructing a power supply change segment sequence, numbering each change segment in sequence and calculating the duration of adjacent segments, a power supply rhythm characteristic sequence is generated, and historical stable operation power supply segments are retrieved to obtain a set of historical stable rhythms. S3. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set to construct a power supply drop prediction window. Generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. S4. When an abnormal candidate segment enters the fall prediction window, obtain the task queue status and execute the beat scheduling process according to the task type. After power is restored, execute task sorting, status verification and task segment merging.

[0021] In this embodiment, S1 aligns the original power supply voltage sequence with a time reference to form a power supply voltage sequence set, and extracts power supply voltage segments by selecting effective operating segments under a unified time coordinate system. Subsequently, S2 generates a power supply rhythm feature sequence by constructing a power supply change segment sequence, sequentially numbering each change segment, and calculating the duration of adjacent segments, and combines this with historical stable operating power supply segments to form a historical stable rhythm set. This process completes the transformation from discrete voltage sampling data to structured rhythm expression, so that power supply fluctuations no longer exist only in the form of instantaneous values, but are organized into a segment structure with sequential relationships and duration characteristics, realizing a continuous characterization of the power supply operating state. Compared with the method of judging only based on single-point voltage thresholds or instantaneous fluctuation amplitude, this scheme constructs a basic reference system by comparing the rhythm feature sequence with the historical stable rhythm set, enabling power supply state analysis to have the ability to compare historical structures. S3 performs sequence alignment comparison processing on the power supply rhythm feature sequence and the historical stable rhythm set to construct a power supply drop prediction window, and generates a task classification table by classifying task metadata according to resource usage patterns, thus establishing a correlation between power supply state identification and task execution structure. This step completes the windowing definition of abnormal candidate segments and the structured identification of task types, enabling the power supply anomaly identification results to be directly mapped to the task scheduling level. Compared with the existing technology where power supply anomalies and task scheduling are handled independently, this solution uses the power supply rhythm identification results as the scheduling trigger condition, realizing pre-judgment of the power supply fluctuation phase rather than post-event response, and organizing preparations in advance at the task execution level. In S4, when an abnormal candidate segment enters the power supply drop prediction window, the rhythm scheduling process is executed according to the task queue status and task type. After power is restored, task sorting, status verification, and task segment merging are performed, forming a complete scheduling closed loop during the power supply fluctuation period and the recovery phase. This processing path completes the task structure rearrangement during the power supply anomaly and the reconstruction of the execution order after recovery, ensuring the continuity and traceability of the task execution process before and after the power supply fluctuation. Compared with the method of simply restarting or restoring tasks uniformly after voltage restoration, this solution, through the drop prediction window and classification scheduling strategy, realizes classified control and recovery reorganization in the power supply fluctuation scenario, enabling the mounted cluster server to have overall control capabilities for structured scheduling and orderly recovery under unstable power supply conditions.

[0022] Example 2 Please refer to Figure 3 Specifically: S1 includes S11; S11. Call the power monitoring interface at the operating system layer of the cluster server computing node to read the node power supply voltage sampling data, form the original power supply voltage sequence, add a timestamp to the original power supply voltage sequence of each node, and then perform time base alignment processing based on the unified time source of the cluster to unify the original power supply voltage sequences of all nodes to the same time coordinate system, and output the power supply voltage sequence set.

[0023] S1 further includes S12; S12. Perform sequence segmentation on the power supply voltage sequence set according to the continuous time window rule, construct several continuous time segments, and perform running status filtering in each time segment, specifically: whether the number of sampling points is lower than the set minimum sampling number, whether the timestamp is in reverse order, and whether the node status log is in restart, shutdown or maintenance state. If any condition is met within each consecutive time segment, the current time segment is marked as an invalid segment. If no condition exists within each consecutive time segment, the current time segment is marked as a valid operating segment, and the time segment that meets the valid operating segment is determined as the power supply voltage segment.

[0024] In this embodiment, in S11, the power monitoring interface is called at the operating system layer of the cluster server computing node to read the node power supply voltage sampling data to form the original power supply voltage sequence. A timestamp is added to the original power supply voltage sequence of each node. Then, based on the unified time source of the cluster, time reference alignment processing is performed to unify the original power supply voltage sequences of all nodes to the same time coordinate system and output the power supply voltage sequence set, so that the power supply data of multiple nodes has a unified time reference basis. Subsequently, in S12, the power supply voltage sequence set is segmented according to the continuous time window rule to construct several continuous time segments. In each time segment, the number of sampling points is checked in turn to see if it is lower than the set minimum sampling number, whether the timestamp is in reverse order, and whether the node status log is in restart, shutdown or maintenance state. Time segments that meet any of the conditions are marked as invalid segments, and time segments that do not meet any of the conditions are marked as valid running segments. The valid running segments are determined as power supply voltage segments. Through the above processing path, the time unification of power supply data, the elimination of abnormal operating states, and the extraction of effective power supply segments were completed. This provides a clear and reliable data foundation for the subsequent construction of power supply change segment sequences, generation of power supply rhythm feature sequences, and construction of power supply drop prediction windows. Compared with the method of directly analyzing the raw voltage data, this avoids the interference of cross-node time deviation and non-operating state data on rhythm judgment, making the power supply rhythm identification results more consistent and comparable, and establishing a stable data premise for the judgment of abnormal candidate segments and task scheduling control.

[0025] Example 3 Please refer to Figure 3 Specifically: S2 includes S21; S21. The voltage sampling sequence in the power supply voltage segment is traversed according to the sampling time order. The voltage values ​​of the current sampling point and the previous sampling point are read. The change direction of the current sampling point is determined according to the voltage magnitude rule between consecutive sampling points. Specifically, when the voltage of the current sampling point is higher than that of the previous sampling point, it is marked as an upward change. When the voltage of the current sampling point is lower than that of the previous sampling point, it is marked as a downward change. When the difference between the two is within the sampling allowable error range, it is marked as a stable change. A continuous segment detection process is performed on the change direction identifier sequence. When the change direction identifiers of adjacent sampling points are consistent, the sampling intervals in the current consistent state are registered as the same change operation segment, and the start and end sampling positions of the current operation segment are recorded. When a change in the change direction identifier of adjacent sampling points is detected, the sampling position where the direction changes is determined as the change boundary position, and multiple change segments within the power supply voltage segment are combined to form a power supply change segment sequence.

[0026] S2 further includes S22; S22. Number each change segment in the power supply change segment sequence according to the order of appearance, and record the duration interval between adjacent segments to generate a power supply rhythm feature sequence. At the same time, retrieve the stable operation power supply segments saved during the historical operation of the node, and generate a historical rhythm feature sequence using the same processing method to obtain a historical stable rhythm set.

[0027] In this embodiment, in S21, the selected power supply voltage segments are traversed point by point according to the sampling time sequence. The voltage values ​​of the current sampling point and the previous sampling point are read. A change direction identifier is generated based on the voltage magnitude relationship between the two. When the difference is within the sampling allowable error range, it is marked as a stable change. Then, continuous segment detection processing is performed on the change direction identifier sequence. Sampling intervals with consistent change directions are registered as the same change operation segment, and the start sampling position and end sampling position are recorded. When the direction changes, the change boundary position is determined, thereby constructing a power supply change segment sequence within a single power supply voltage segment. Further, in S22, the power supply change segment sequence is numbered according to the order of appearance, and the duration interval between adjacent segments is recorded to form a power supply rhythm feature sequence containing segment sequence and duration characteristics. At the same time, historical stable operation power supply segments are retrieved and historical rhythm feature sequences are generated using the same processing method to construct a historical stable rhythm set. Through the above implementation path, the original continuous voltage waveform is transformed into a structured rhythmic expression with directional, segmental, and time span attributes, so that the power supply status has numberable, alignable, and comparable sequence characteristics. This provides a unified structural basis for the subsequent sequence alignment and comparison of the power supply rhythmic feature sequence and the historical stable rhythm set, achieving the purpose of structurally characterizing the power supply operation rhythm and establishing historical references. This enables the subsequent identification of abnormal candidate segments and the construction of the power supply drop prediction window to be based on the ordered segment model.

[0028] Example 4 Please refer to Figure 3 Specifically: S3 includes S31; S31. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set, as follows; When the arrangement of the changing segments in the power supply rhythm characteristic sequence is consistent with the historical stable rhythm set, the corresponding power supply voltage segment is marked as a stable power supply segment. The voltage data in each change segment of the power supply rhythm characteristic sequence is compared point by point. When the voltage of multiple sampling points shows a monotonically decreasing trend and the decrease is greater than the set decrease threshold, the current change segment is marked as a decreasing change segment. When the voltage data after the decreasing change segment reverses and rises, and the rise is greater than the set rise threshold, the current change segment is marked as a recovery change segment. When a combination of decreasing and recovering change segments is detected in the current power supply rhythm characteristic sequence, and the current combination does not appear in the historical stable rhythm set, the corresponding power supply voltage segment is identified as an abnormal candidate segment. The starting point of the abnormal candidate segment combination is used as the window start point, and the position where the change recovery is completed is used as the window end point to generate a power supply drop prediction window. The threshold for the decrease is set based on the maximum decrease in normal power supply fluctuations in historical data, and the threshold for the recovery is set based on the minimum effective recovery amplitude of the voltage recovery process in historical data.

[0029] S3 further includes S32; S32. Obtain metadata of tasks to be executed and tasks currently being executed within the mounted cluster server through the Kubernetes scheduler. The metadata includes task type, resource consumption, execution time, CPU usage and I / O frequency. Use a classification algorithm to classify the resource usage patterns of the metadata into continuous occupation type, burst type, synchronous sensitive type and segmentable type, and generate a task classification table containing a unique identifier, task type, required resources and lifecycle data for each task. Sustained-occupancy type: The task consumes CPU and memory computing resources steadily throughout the entire execution process, and there is no fluctuation in resource consumption; Impulse burst type: High-concurrency network requests or disk read / write operations within a short period of time; Synchronization-sensitive: The task requires frequent synchronization operations between multiple nodes, and the synchronization operations have time requirements; Segmentable: The task can be split into several short execution segments, and each execution segment runs independently. Pauses, migrations or other optimization strategies can be inserted between segments.

[0030] In this embodiment, firstly, S31 performs sequence alignment and comparison processing on the power supply rhythm feature sequence and the historical stable rhythm set. When the arrangement of the change segments is consistent, it is determined to be a stable power supply segment. When a decreasing change segment that meets the decrease amplitude threshold and a recovering change segment that meets the recovery amplitude threshold are identified in the power supply rhythm feature sequence, and the combination of the decreasing change segment and the recovering change segment does not appear in the historical stable rhythm set, the corresponding power supply voltage segment is determined as an abnormal candidate segment. A power supply drop prediction window is constructed based on the starting boundary point of the abnormal candidate segment combination and the position where the change recovery is completed. On this basis, S32 obtains the metadata of the tasks to be executed and the tasks being executed through the Kubernetes scheduler, and classifies them according to the resource usage mode as continuous occupation type, burst type, synchronization sensitive type, and segmentable type. The system generates a task classification table containing a unique task identifier, task type, required resources, and lifecycle data, thus establishing a correspondence between power supply rhythm identification results and task execution structure. Through the above implementation method, power supply anomaly identification is transformed from single-point voltage judgment to a combined discrimination based on the structure of changing segments and historical stable rhythm sets. A task type structure mapping relationship is established within the power supply drop prediction window, creating a pre-linkage path between power supply fluctuations and task scheduling. This achieves the goal of identifying abnormal candidate segments in advance and completing task structure classification preparation during the power supply fluctuation stage. Compared with the method of separating power supply anomaly and task scheduling, this enables the mounted cluster server to have the ability to predict based on rhythm structure and the basis for classification scheduling in power supply unstable scenarios, forming a structured connection effect between power supply anomaly identification and task execution organization.

[0031] Example 5 Please refer to Figure 3 Specifically: S4 includes S41; S41. When an abnormal candidate segment of a node enters the fall prediction window, extract the task queue status from the current node's task classification table and start the beat scheduling process, as follows; To identify burst-type tasks, a parallelism convergence algorithm is used to merge multiple parallel tasks, and a batch processing algorithm is used to merge several tasks into a single task. Identify segmentable tasks and preset task segmentation thresholds based on the shortest execution task during historical stable operation. When each execution segment in a segmentable task exceeds the task segmentation threshold, a task splitting algorithm is used to split it into several sub-execution segments. The length of each execution segment is dynamically adjusted according to the power supply fluctuation cycle. Identify synchronization-sensitive tasks and use a dynamic synchronization-triggered scheduling algorithm to stagger cross-node tasks. Identify persistently occupied tasks, freeze the current task using a protective freeze algorithm, and resume when voltage fluctuations return to stability.

[0032] S4 further includes S42; S42. After the power supply is restored to a stable state, tasks are sorted and scheduled based on the historical resource scheduling analysis algorithm. When a task that needs to be synchronized across nodes is detected, the synchronization alignment point of the task is restored through the timestamp synchronization mechanism. The task segments are checked one by one according to the log analysis algorithm to confirm the execution status of each task segment. The task segments are then merged according to the execution order and timestamp of the tasks.

[0033] In this embodiment, when an abnormal candidate segment of a node enters the drop prediction window, S41 extracts the current task queue status based on the task classification table, starts the tick scheduling process, and performs structured processing on tasks with different resource usage modes: for burst-type tasks, concurrent compression and request frequency constraints are performed through parallel convergence algorithms and batch processing algorithms to ensure that high-frequency instantaneous loads present orderly output within the power supply fluctuation cycle; for segmentable tasks, a task splitting algorithm is performed based on the shortest execution task preset task segmentation threshold during historical stable operation to match the execution segment length with the power supply fluctuation cycle, so that the computational load corresponds to the power supply rhythm in the time dimension; For synchronization-sensitive tasks, a dynamic synchronization triggering scheduling algorithm is used to adjust the cross-node synchronization timing, reducing the synchronization concentration within the power supply fluctuation window. For tasks with continuous occupancy, a protective freeze algorithm is executed, suspending their resource occupation during voltage fluctuations and resuming execution when the voltage returns to stability. Subsequently, after the power supply returns to a stable state, S42 sorts tasks using a historical resource scheduling analysis algorithm. When cross-node synchronization tasks exist, a timestamp synchronization mechanism is used to restore the synchronization alignment point. Combined with a log analysis algorithm, task segments are checked one by one and merged according to execution order and timestamp, so that the reorganized task structure within the window period returns to a unified execution trajectory. Through the above process, a correspondence is established between the power supply drop prediction window and the task scheduling strategy. During power supply fluctuations, the task load rhythm is rearranged, and during the power supply recovery phase, the execution sequence is reorganized. This ensures that the mounted cluster server maintains continuous operation before and after power supply anomalies, preventing power supply fluctuations from directly transmitting to large-scale task impacts, and achieving the goal of categorized control and recovery reconstruction driven by power supply rhythm perception.

[0034] Example 6 Please refer to Figure 2 A control system for mounted cluster servers includes a power supply rhythm acquisition module, an anomaly identification module, a drop prediction module, and a fingerprint database write-back module. The power supply rhythm acquisition module is used to call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment processing, a power supply voltage sequence set is formed, and the effective operating segment is selected to extract the power supply voltage segment. The anomaly identification module is used to construct a power supply change segment sequence, number each change segment sequentially and calculate the duration of adjacent segments, generate a power supply rhythm feature sequence, and retrieve historical stable operation power supply segments to obtain a set of historical stable rhythms. The power supply drop prediction module is used to perform sequence alignment and comparison processing on the power supply rhythm feature sequence and the historical stable rhythm set, construct the power supply drop prediction window, and generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. The fingerprint database write-back module is used to obtain the task queue status and execute the clock scheduling process according to the task type when abnormal candidate segments enter the fall prediction window. After power is restored, it performs task sorting, status verification and task segment merging.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A control method for mounted cluster servers, characterized in that: Includes the following steps: S1. Call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment, form a power supply voltage sequence set and filter the valid operating segments to extract the power supply voltage segments. S2. By constructing a power supply change segment sequence, numbering each change segment in sequence and calculating the duration of adjacent segments, a power supply rhythm characteristic sequence is generated, and historical stable operation power supply segments are retrieved to obtain a set of historical stable rhythms. S3. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set to construct a power supply drop prediction window. Generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. S4. When an abnormal candidate segment enters the fall prediction window, obtain the task queue status and execute the beat scheduling process according to the task type. After power is restored, execute task sorting, status verification and task segment merging.

2. The control method for a mounted cluster server according to claim 1, characterized in that: S1 includes S11; S11. Call the power monitoring interface at the operating system layer of the cluster server computing node to read the node power supply voltage sampling data, form the original power supply voltage sequence, add a timestamp to the original power supply voltage sequence of each node, and then perform time base alignment processing based on the unified time source of the cluster to unify the original power supply voltage sequences of all nodes to the same time coordinate system, and output the power supply voltage sequence set.

3. The control method for a mounted cluster server according to claim 2, characterized in that: S1 further includes S12; S12. Perform sequence segmentation on the power supply voltage sequence set according to the continuous time window rule, construct several continuous time segments, and perform running status filtering in each time segment, specifically: whether the number of sampling points is lower than the set minimum sampling number, whether the timestamp is in reverse order, and whether the node status log is in restart, shutdown or maintenance state. If any condition is met within each consecutive time segment, the current time segment is marked as an invalid segment. If no condition exists within each consecutive time segment, the current time segment is marked as a valid operating segment, and the time segment that meets the valid operating segment is determined as the power supply voltage segment.

4. The control method for a mounted cluster server according to claim 3, characterized in that: S2 includes S21; S21. The voltage sampling sequence in the power supply voltage segment is traversed according to the sampling time order. The voltage values ​​of the current sampling point and the previous sampling point are read. The change direction of the current sampling point is determined according to the voltage magnitude rule between consecutive sampling points. Specifically, when the voltage of the current sampling point is higher than that of the previous sampling point, it is marked as an upward change. When the voltage of the current sampling point is lower than that of the previous sampling point, it is marked as a downward change. When the difference between the two is within the sampling allowable error range, it is marked as a stable change. A continuous segment detection process is performed on the change direction identifier sequence. When the change direction identifiers of adjacent sampling points are consistent, the sampling intervals in the current consistent state are registered as the same change operation segment, and the start and end sampling positions of the current operation segment are recorded. When a change in the change direction identifier of adjacent sampling points is detected, the sampling position where the direction changes is determined as the change boundary position, and multiple change segments within the power supply voltage segment are combined to form a power supply change segment sequence.

5. The control method for a mounted cluster server according to claim 4, characterized in that: S2 further includes S22; S22. Number each change segment in the power supply change segment sequence according to the order of appearance, and record the duration interval between adjacent segments to generate a power supply rhythm feature sequence. At the same time, retrieve the stable operation power supply segments saved during the historical operation of the node, and generate a historical rhythm feature sequence using the same processing method to obtain a historical stable rhythm set.

6. The control method for a mounted cluster server according to claim 5, characterized in that: S3 includes S31; S31. Perform sequence alignment and comparison processing on the power supply rhythm characteristic sequence and the historical stable rhythm set, as follows; When the arrangement of the changing segments in the power supply rhythm characteristic sequence is consistent with the historical stable rhythm set, the corresponding power supply voltage segment is marked as a stable power supply segment. The voltage data in each change segment of the power supply rhythm characteristic sequence is compared point by point. When the voltage of multiple sampling points shows a monotonically decreasing trend and the decrease is greater than the set decrease threshold, the current change segment is marked as a decreasing change segment. When the voltage data after the decreasing change segment reverses and rises, and the rise is greater than the set rise threshold, the current change segment is marked as a recovery change segment. When a combination of decreasing and recovering change segments is detected in the current power supply rhythm characteristic sequence, and the current combination does not appear in the historical stable rhythm set, the corresponding power supply voltage segment is identified as an abnormal candidate segment. The starting point of the abnormal candidate segment combination is used as the window start point, and the position where the change recovery is completed is used as the window end point to generate a power supply drop prediction window. The threshold for the decrease is set based on the maximum decrease in normal power supply fluctuations in historical data, and the threshold for the recovery is set based on the minimum effective recovery amplitude of the voltage recovery process in historical data.

7. The control method for a mounted cluster server according to claim 6, characterized in that: S3 further includes S32; S32. Obtain metadata of tasks to be executed and tasks currently being executed within the onboard cluster server through the Kubernetes scheduler. The metadata includes task type, resource consumption, execution time, CPU usage, and I / O frequency. Use a classification algorithm to classify the resource usage patterns of the metadata into continuous occupation type, burst type, synchronous sensitive type, and segmentable type, and generate a task classification table containing a unique identifier, task type, required resources, and lifecycle data for each task.

8. The control method for a mounted cluster server according to claim 7, characterized in that: S4 includes S41; S41. When an abnormal candidate segment of a node enters the fall prediction window, extract the task queue status from the current node's task classification table and start the beat scheduling process, as follows; To identify burst-type tasks, a parallelism convergence algorithm is used to merge multiple parallel tasks, and a batch processing algorithm is used to merge several tasks into a single task. Identify segmentable tasks and preset task segmentation thresholds based on the shortest execution task during historical stable operation. When each execution segment in a segmentable task exceeds the task segmentation threshold, a task splitting algorithm is used to split it into several sub-execution segments. The length of each execution segment is dynamically adjusted according to the power supply fluctuation cycle. Identify synchronization-sensitive tasks and use a dynamic synchronization-triggered scheduling algorithm to stagger cross-node tasks. Identify persistently occupied tasks, freeze the current task using a protective freeze algorithm, and resume when voltage fluctuations return to stability.

9. A control method for a mounted cluster server according to claim 8, characterized in that: S4 further includes S42; S42. After the power supply is restored to a stable state, tasks are sorted and scheduled based on the historical resource scheduling analysis algorithm. When a task that needs to be synchronized across nodes is detected, the synchronization alignment point of the task is restored through the timestamp synchronization mechanism. The task segments are checked one by one according to the log analysis algorithm to confirm the execution status of each task segment. The task segments are then merged according to the execution order and timestamp of the tasks.

10. A control system for a mounted cluster server, comprising the control method for a mounted cluster server as described in any one of claims 1-9, characterized in that: It includes a power supply rhythm acquisition module, an anomaly identification module, a drop prediction module, and a fingerprint database write-back module; The power supply rhythm acquisition module is used to call the power monitoring interface to read the original power supply voltage sequence and add a timestamp. After time reference alignment processing, a power supply voltage sequence set is formed, and the effective operating segment is selected to extract the power supply voltage segment. The anomaly identification module is used to construct a power supply change segment sequence, number each change segment sequentially and calculate the duration of adjacent segments, generate a power supply rhythm feature sequence, and retrieve historical stable operation power supply segments to obtain a set of historical stable rhythms. The power supply drop prediction module is used to perform sequence alignment and comparison processing on the power supply rhythm feature sequence and the historical stable rhythm set, construct the power supply drop prediction window, and generate a task classification table by acquiring task metadata and classifying it according to resource usage mode. The fingerprint database write-back module is used to obtain the task queue status and execute the clock scheduling process according to the task type when abnormal candidate segments enter the fall prediction window. After power is restored, it performs task sorting, status verification and task segment merging.

Images