A computing power resource dynamic scheduling system fusing environment perception and power self-balancing

Abstract

The application discloses a kind of computing power resource dynamic scheduling systems of fusion environment perception and power self-balancing, it is related to computing power resource scheduling technical field.The system, through data acquisition and environment perception fusion module, real-time acquisition is carried out in the multi-dimensional data of the specified computing power facility inside and outside of server cluster operation, and with current environment perception parameter dynamic / static fusion processing, realize the comprehensive perception of environment and operating state;Then through joint scheduling and self-power balance control module combination fusion processing result optimization resource allocation;Finally, through layered scheduling and multi-scenario adaptation verification module generates layered scheduling scheme, combined with progressive derating determination and fault tolerance rate check, determine whether the layered scheduling scheme generated is adapted to the current scene, realizes the dynamic linkage regulation and control of load, environment and power, and effectively solves the problem of power distribution imbalance in computing power scheduling, improves the energy efficiency ratio and environmental adaptability of server cluster.

Description

A dynamic scheduling system for computing resources that integrates environmental perception and power self-balancing Technical Field

[0001] This invention relates to the field of computing resource scheduling technology, and in particular to a dynamic computing resource scheduling system that integrates environmental awareness and power self-balancing. Background Technology

[0002] With the major trends of enterprise digital transformation and the industrial application of artificial intelligence, the requirements for the efficiency, stability, and energy efficiency of computing infrastructure (such as data centers and server rooms) supporting various business operations are becoming increasingly stringent. Data centers are the core computing power carriers that centrally deploy servers, storage, and network equipment. They are divided into large-scale Internet-level (carrying cloud services and big data processing) and enterprise-level (supporting internal business), characterized by high-density computing power and 24 / 7 uninterrupted operation. Computing infrastructure refers to small-scale computing scenarios built by enterprises for local data processing and business support.

[0003] Computing resource scheduling is directly related to power consumption. High loads can lead to scheduling lags causing a surge in power consumption, while low loads result in idle resources and wasted power. For example, as the scale of server clusters supported by these two types of facilities expands, and the demand for computing power from services such as cloud computing surges, computing load fluctuations intensify, highlighting the pressure on energy consumption and power management. Simultaneously, they are susceptible to temperature and humidity: high temperatures or heat dissipation failures in summer can reduce server cooling efficiency, forcing the scheduling system to reduce frequency and limit current; high humidity can cause short circuits, and dryness can generate static electricity, both of which interfere with the stability of computing resource scheduling; high-density cluster operation in large data centers easily generates heat, and poor ventilation can cause localized hotspots, further exacerbating power loss and scheduling difficulties. Therefore, coordinated management of computing resource scheduling and power is necessary to achieve efficient and stable facility operation.

[0004] For example, Chinese Invention Patent No. CN119248490B discloses an intelligent computing power scheduling method and system based on dynamic programming, which includes: assessing the balance of resource allocation by real-time monitoring of the current available resource information of multiple computing power clusters and calculating the resource difference value between clusters to determine a preliminary computing power grouping scheme; based on this, matching suitable task types for each group according to historical task types and resource consumption characteristics, and adjusting the preliminary grouping scheme to form an optimized computing power grouping; using the optimized computing power grouping, calculating the service capability score of each group according to a preset service level agreement, and providing users with a recommended set of computing power resource services after sorting by score.

[0005] For example, the computing resource scheduling method, computing resource scheduling device, and computer equipment disclosed in Chinese invention patent application CN120596277A include: collecting computing resource data and task requirement data, synchronously generating resource nodes and resource node attributes, and constructing a heterogeneous network; in the heterogeneous network, performing random walks between nodes according to set random walk rules to generate multiple random walk paths, and using the random walk paths as training text to map the words in them to corresponding resource node vectors and task node vectors; inputting the obtained task-resource combinations into a multilayer perceptron model, and calculating the scores of each group through the multilayer perceptron model; and scheduling the resource with the highest score obtained after combining with the current task as the matching resource.

[0006] In summary, existing technologies primarily focus on the balanced allocation of computing resources and the optimization of task-resource matching. Specifically, they dynamically allocate computing resources within the infrastructure by real-time collection of server hardware status (e.g., CPU / GPU load, memory usage), the physical environment of the computing infrastructure (e.g., rack temperature, air conditioning operation status), business load (e.g., task type, real-time requirements), and energy supply (e.g., power load of the computing infrastructure, green electricity access). Simultaneously, some solutions attempt to correlate the environment and operational status with fundamental responses. For example, by setting a fixed power limit per rack in the computing infrastructure and allocating power quotas based on the server's computing power percentage, they achieve computing power scheduling and power control within the infrastructure, ensuring stable operation.

[0007] Furthermore, existing technologies do not fully consider the coupling relationship between the environment and operating status. Specifically, in terms of coupling response, some computing infrastructure directly allocates power quotas according to fixed standards or only allocates power resources based on the proportion of server hardware configuration, without fully considering load fluctuations and environmental differences during actual operation. This results in a lack of flexibility in power allocation from the outset. Moreover, during computing power scheduling, existing solutions do not deeply correlate real-time load data with power adjustment during the environmental coupling response process. For example, when servers are under high load, real-time load data is not promptly analyzed in conjunction with environmental parameters, causing data processing delays and preventing power adjustment commands from keeping pace with load changes. Conversely, servers under low load, lacking dynamic linkage logic between load and power, maintain a fixed base power consumption and cannot adjust power synchronously as the load decreases. This prevents power adjustment from predictively adjusting to environmental changes, leading to insufficient adaptability between environmental adaptability and server cluster operating energy consumption during computing power scheduling, resulting in uneven power allocation in dynamic scheduling of computing resources. Summary of the Invention

[0008] To address the technical problem of uneven power allocation in the dynamic scheduling of computing resources in existing technologies, this invention provides a dynamic scheduling system for computing resources that integrates environmental awareness and power self-balancing. The technical solution is as follows: a data acquisition and environmental awareness fusion module, used to collect multi-dimensional data from inside and outside the designated computing facility where the server cluster operates in real time through a sensor network, and fuse this data with current environmental awareness parameters; a joint scheduling and self-power balancing control module, used to quantitatively determine joint scheduling based on the fusion processing results, and dynamically adjust the power quota of each server in the server cluster based on the acquired power consumption data; and a hierarchical scheduling and multi-scenario adaptation verification module, used to generate a hierarchical scheduling scheme based on the power quota constraints, and to determine gradual and collaborative derating, while also verifying the fault tolerance rate of the derating process.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0010] 1. This invention utilizes a data acquisition and environmental perception fusion module with a sensor network to collect multi-dimensional data in real time, such as temperature and humidity of computing facilities, server load rate, and airflow cooling capacity. This data is deeply integrated with the perceived parameters of the current environment, avoiding power adjustment inaccuracies caused by data lag. The joint scheduling and self-power balancing control module performs joint scheduling based on the fusion results, optimizing resource allocation efficiency. At the same time, it uses a power self-balancing mechanism to dynamically constrain quotas, effectively addressing load fluctuations and environmental differences, and maintaining a dynamic balance between power fluctuations and performance requirements. The hierarchical scheduling and multi-scenario adaptation verification module generates hierarchical scheduling schemes based on power quota constraints. Combined with progressive derating judgment and fault tolerance verification, it ensures that the scheme is accurately adapted to the current scheduling scenario. This solves the problem of power not keeping up with the time under high load and avoids ineffective power consumption under low load. Ultimately, it achieves deep linkage between load, environment, and power, effectively improving the energy efficiency ratio and environmental adaptability of the server cluster. This allows power adjustment to make predictive adjustments in response to environmental changes, optimizing operating energy consumption and adaptability.

[0011] 2. Dynamic coupling processing is performed across different spatiotemporal dimensions: First, multi-dimensional data is divided into environment and computing power dimensions. After standardization and spatiotemporal alignment, corresponding feature vectors are obtained. The former reflects the dynamic distribution of the environment, while the latter reflects the real-time changes in cluster operation. Then, a nonlinear mapping is performed using an environment-operation coupling function aimed at minimizing the impact of the environment on performance, outputting an environment-operation coupling correlation coefficient that quantifies the adaptability. Subsequently, the first derivative of this coefficient is performed to obtain the instantaneous rate of change. Finally, the results are concatenated to generate the dynamic coupling processing result, which can accurately capture the real-time collaborative relationship between the environment and the operating state, providing timely basis for dynamic scheduling. Static coupling processing focuses on the same spatiotemporal dimension: After normalization and static correlation calibration, multi-dimensional data yields an environment benchmark vector reflecting the steady-state distribution of the environment and a computing power benchmark vector reflecting the benchmark state of the cluster. A linear mapping is performed using an environment-operation static matching function aimed at maximizing parameter matching degree, outputting an environment-operation static matching coefficient that quantifies the benchmark adaptability level. After obtaining the cumulative change through second derivative, the results are superimposed to generate the static coupling processing result, which can stably reflect the benchmark collaborative relationship between the environment and operation, avoiding short-term fluctuations from interfering with decision-making. The entire process, through differentiated dimensional processing, targeted function calculations, and multi-dimensional data fusion, not only achieves real-time dynamic perception of the environment and operating status, but also ensures stable determination of baseline adaptability, effectively improving the comprehensiveness and accuracy of data fusion, and ensuring that scheduling decisions are more in line with actual operational needs.

[0012] 3. The quantitative judgment of joint scheduling is classified into two categories based on dynamic and static coupling processing results: If the result is dynamic coupling processing, the CPU / GPU base frequency is preferentially triggered for rapid adjustment and cross-node load migration is prepared simultaneously to ensure rapid matching of high load demand; if the limit is not exceeded, the frequency adjustment gradient is maintained, and only migration conditions are monitored. If the node load difference exceeds the limit, migration is executed; otherwise, monitoring continues. In the migration condition judgment, if the environment-operation coupling correlation coefficient is lower than the limit, migration is paused and the frequency is lowered to the lower limit of the interval to ensure operational safety and effectively cope with load fluctuations and dynamic changes in the environment. If the result is static coupling processing, the current frequency is maintained first. If the node load difference exceeds the static limit, resources are allocated according to priority to execute migration to avoid excessive load deviation between nodes; if the limit is not exceeded, the cumulative change is monitored. If the limit is exceeded, a trend imbalance warning is issued; otherwise, the operating parameters are maintained to ensure load balance under steady state. The entire judgment process, through differentiated control logic, achieves both rapid response and safety net in dynamic scenarios, and ensures steady-state equilibrium and risk warning in static scenarios. This effectively improves the pertinence and reliability of scheduling decisions. At the same time, through clear limit judgments and gradient adjustments, it avoids power waste and performance loss, ensuring the stable and efficient operation of the server cluster.

[0013] 4. By normalizing the current power consumption data, the median power consumption is calculated using the median sorting method. Then, the deviation of each node's real-time power from the median is calculated. If the absolute value of a node's deviation exceeds a preset proportion, a first-level adjustment is triggered, adjusting its power quota in a step-by-step manner according to the excess range. This can quickly correct power imbalances in a single node. If the power of a single node still exceeds the preset median proportion during the first-level adjustment, a second-level adjustment is triggered, increasing the quota adjustment range on top of the original step-by-step adjustment. For load migration scenarios, if the power of a target node exceeds the preset median proportion after migration, its quota is temporarily reduced to avoid new imbalances caused by migration and ensure cluster power stability. If the total power still exceeds the rated value after the first-level adjustment, an anomaly alert is sent, prioritizing the safe operation of the cluster. The entire process, through logic such as quantifying deviations, tiered adjustments, and migration constraints, achieves dynamic and accurate correction of power quotas. This not only solves the problem of abnormal power in a single node but also avoids secondary imbalances caused by load migration, effectively improving the balance and operational stability of server cluster power allocation.

[0014] 5. By comparing the current operating data of the hierarchical scheduling scheme with the reference operating data, if the energy efficiency ratio is lower than the preset proportion of the reference value or the power fluctuation exceeds the range, it is determined to be a gradual derating; otherwise, it is a coordinated derating. The derating requirement is clarified through data quantification and comparison, avoiding blind regulation. Gradual derating is executed in a gradient according to the hierarchical order, while the number of task failures at each layer is counted. Coordinated derating is based on the total power target of the cluster, and the derating of each layer is reduced proportionally, while the number of cross-layer data interaction failures is counted. The differentiated derating logic can be adapted to scenarios with insufficient energy efficiency, excessive power fluctuations, and conventional power optimization, improving the targeting of regulation. In the gradual throttling process, if the number of task failures at any layer exceeds the corresponding allowable value, the process at that layer is paused and a prompt is displayed to re-verify the power quota. In the collaborative throttling process, if the number of cross-layer interaction failures exceeds the limit, the entire solution is paused and a prompt is displayed to adjust the quota allocation ratio before re-verification. If no pause operation is triggered during the throttling process, the solution is determined to be suitable for the current scenario and is continuously monitored. This ensures the accuracy and effectiveness of the throttling while minimizing scheduling risks and ensuring that the server cluster maintains a stable operating state while optimizing energy consumption and power balance. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 is a schematic diagram of the structure of a dynamic scheduling system for computing resources that integrates environmental perception and power self-balancing according to an embodiment of the present invention;

[0017] Figure 2 is a flowchart of the coupling process provided in an embodiment of the present invention;

[0018] Figure 3 is a flowchart of the quantitative determination process for joint scheduling provided in an embodiment of the present invention;

[0019] Figure 4 is a flowchart of the dynamic correction process of power quotas for each server in a server cluster provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] This invention provides a dynamic scheduling system for computing resources that integrates environmental awareness and power self-balancing. Figure 1 shows a schematic diagram of the structure of such a system. The system includes:

[0024] The data acquisition and environmental perception fusion module is used to collect multi-dimensional data from inside and outside the designated computing facility where the server cluster operates in real time through a sensor network, and fuse it with current environmental perception parameters to achieve comprehensive perception of the environment and operating status of the server cluster. The multi-dimensional data reflects the environmental conditions inside and outside the designated computing facility and the operating status of the server cluster. This multi-dimensional data typically includes the temperature and humidity of the designated computing facility, server load rate, airflow cooling capacity, power grid fluctuations, and heat distribution between server racks. The joint scheduling and self-power balancing control module is used to perform joint scheduling based on the fusion processing results to optimize the resource allocation efficiency of the server cluster, while also introducing power... A power self-balancing mechanism is implemented to constrain power quotas and maintain a dynamic balance between power fluctuations and performance requirements of the server cluster. A hierarchical scheduling and multi-scenario adaptation verification module is used to generate a hierarchical scheduling scheme based on the power quota constraints, while performing progressive derating judgment and fault tolerance verification to verify the adaptability of the hierarchical scheduling scheme to the current scheduling scenario of the server cluster, thereby effectively improving energy efficiency and environmental adaptability. The hierarchical scheduling scheme includes a redundant resource scheduling layer, an edge load balancing scheduling layer, and a task scheduling layer. The power quota allocation and computing load process corresponding to the three scheduling layers are carried out progressively, that is, according to the principle of progressive resource priority and power consumption adaptation.

[0025] Specifically, multi-dimensional data is acquired collaboratively through distributed sensors and monitoring terminals deployed throughout the computing facility: temperature and humidity and heat distribution between racks are captured in real time by temperature sensors and thermal imaging equipment deployed throughout the facility; airflow cooling capacity is collected jointly by wind speed sensors and pressure monitoring devices within the airflow duct; power grid power fluctuations rely on power monitoring terminals to record voltage and current fluctuation parameters; and server load rate is obtained directly by reading hardware operating indicators.

[0026] The joint scheduling and self-power balancing control module includes a joint scheduling and judgment unit and a self-power balancing control unit. The joint scheduling and judgment unit receives the coupling processing results and, based on the current load distribution and task priority requirements of the server cluster, performs quantitative judgment on joint scheduling to output joint scheduling instructions that meet performance constraints. The coupling processing results include dynamic and static coupling processing results. Joint scheduling includes dynamic adjustment of the CPU / GPU base frequency to match real-time load fluctuations and adapt to dynamic environmental changes, and cross-node load migration to balance node load pressure and avoid local overheating. The CPU / GPU base frequency represents the initial operating frequency of the server cluster set by preset personnel in the current operating state. The self-power balancing control unit dynamically corrects the power quota of each server in the server cluster based on the output joint scheduling instructions and real-time power consumption data, under the closed-loop feedback control logic of the server cluster, ensuring a dynamic balance between performance output and power consumption. The power consumption data reflects the dynamic changes in the real-time operating power consumption of each server in the server cluster, the peak power of a single node, and the total energy consumption of the entire cluster.

[0027] In a specific implementation of large-scale data center server cluster scheduling, the system achieves precise scheduling through the collaborative efforts of three modules: First, the data acquisition and environmental perception fusion module relies on distributed temperature sensors and thermal imaging equipment throughout the computing facility to capture multi-dimensional data such as temperature, humidity, and heat distribution between racks in real time. This data is then fused with environmental parameters to generate dynamic and static coupled results, providing comprehensive and accurate perception data for subsequent scheduling. Next, the joint scheduling and judgment unit, combining the cluster's real-time load distribution and task priorities, outputs CPU / GPU base frequency adjustment and cross-node load migration instructions. The self-power balancing control unit, through closed-loop feedback, dynamically corrects quotas for single-node power deviations exceeding limits, effectively avoiding load imbalance and power waste, and maintaining a dynamic balance between performance and power consumption. Finally, the hierarchical scheduling module generates hierarchical schemes based on quota constraints. Through progressive derating judgment and fault tolerance verification, it adapts to scenario requirements, ensuring continuous operation of core tasks while improving the system's adaptability to complex environments. The entire process, through the closed-loop logic of perception, scheduling, and verification, significantly improves the energy efficiency ratio, operational stability, and scenario adaptability of the server cluster.

[0028] Since existing technologies mostly rely on static modes for environmental coupling responses, resulting in insufficient dynamic correlation and an inability to adapt to complex changes in computing load and environment, a coupling processing flowchart as shown in Figure 2 is provided. This flowchart employs differentiated spatiotemporal dimensions and mapping algorithms, using both dynamic and static dual-path coupling processing to achieve precise quantitative correlation between the environment and computing power operating status. The specific coupling processing process is as follows:

[0029] In Example 1, since the computing facility environment and the server cluster operating status are dynamically correlated in different times and spaces, the specific process of fusion processing in different time and space dimensions includes: classifying the collected multi-dimensional data into data types to obtain environmental dimension data (i.e., temperature and humidity, air duct heat dissipation capacity, power grid fluctuations, and heat distribution between racks) and computing power dimension data (i.e., server cluster operating status parameters, with server load rate as the core). Standardization processing is performed on the two types of data to eliminate the differences in the dimensions and numerical magnitudes of different indicators, ensuring data comparability; at the same time, time and space alignment processing is performed to uniformly map data from different collection times and different deployment locations to a consistent time and space coordinate system, correcting time lag and spatial misalignment problems, and finally forming a well-structured environmental feature vector with data from the same source. and computing power feature vector The input is then fed into the constructed environment-run coupling function for nonlinear mapping operations. This function is implemented using radial basis functions, and the calculation formula is as follows: In the formula, The Euclidean distance between the environmental feature vector and the computing power feature vector is used to quantify the spatial difference between the two types of features. This represents the kernel width, used to adjust the sensitivity of the nonlinear mapping. This represents the weight coefficients obtained through historical training, used to adapt the association weights between the environment and the running state under different scenarios, and thus output the environment-running coupling correlation coefficient. .

[0030] Subsequently, the environment-operation coupling correlation coefficient was subjected to first-order differential processing in the continuous spatiotemporal dimension, and the instantaneous rate of change was solved using the central difference method. The calculation formula is: In the formula, t represents the time variable. The spatiotemporal sampling interval (consistent with the data acquisition frequency) is used to obtain the instantaneous rate of change of environmental parameters and operational status response. Finally, the obtained instantaneous rate of change is concatenated with the environmental feature vector and the computing power feature vector according to the formula. To generate real-time feature vectors that reflect the collaborative relationship between the environment and operational status. The result is denoted as the result of dynamic coupling processing. Here, T represents the vector transpose operation, which is used to convert the environmental feature vector and the computing power feature vector from column vectors to row vectors to meet the horizontal format requirements of dimension splicing processing. The matrix formed by splicing the instantaneous rate of change with the environmental feature vector and the computing power feature vector is transposed as a whole, so that the final generated real-time feature vector maintains a unified format in dimensional structure, thereby ensuring the integrity and consistency of the feature vector.

[0031] Among them, the environmental feature vector is used to reflect the dynamic distribution of the environmental perception parameters of the specified computing power facility in different areas and under different operating periods; the computing power feature vector is used to reflect the real-time changes of the corresponding operating equipment of the server cluster under different computing power load pressures; the constructed environment-operation coupling function represents the objective function of minimizing the impact of environmental fluctuations on equipment performance, and is used to quantify the dynamic correlation between the environmental perception parameters of the specified computing power facility and the operating status of the server cluster; the environment-operation coupling correlation coefficient is used to quantify the degree of influence of the environmental perception parameters in the dynamic environment of the specified computing power facility on the operating status of the server cluster, that is, the adaptability between the environmental perception parameters and the operating status of the server cluster.

[0032] In this embodiment, the fusion process accurately captures the dynamic correlation between the computing facility environment and the server cluster's operating status. Through data type segmentation and standardization, and spatiotemporal alignment, it ensures the consistency and comparability of environment and computing power dimension data, laying a high-quality data foundation for subsequent coupling operations. Based on the environment-operation coupling function constructed using radial basis functions, the dynamic adaptability between the environment and operating status is accurately quantified with the goal of minimizing the environment's impact on performance. The instantaneous rate of change obtained through first-order differential processing captures the dynamic trends of environmental changes and operational responses in real time. Combined with the dynamic coupling results generated by dimension splicing, it comprehensively integrates the characteristics of the environment, computing power, and rate of change. The overall process achieves deep fusion and accurate representation of the environment and operating status, effectively improving the targeting and scientific nature of scheduling decisions and ensuring the stable and efficient operation of the server cluster in a dynamic environment.

[0033] Based on Example 1, since computing facilities are not always in a dynamic fluctuation state during operation, there are scenarios where environmental parameters (such as temperature and humidity, power grid supply) and server cluster operating status (such as load rate, hardware utilization rate) are relatively stable. Under the same spatiotemporal dimension, the two exhibit a more static correlation at the baseline level. Therefore, under the same spatiotemporal dimension, the specific process of fusion processing is as follows:

[0034] The collected multi-dimensional data were normalized and subjected to static correlation calibration. Normalization employed the min-max standardization method to eliminate differences in data magnitude. Static correlation calibration corrected system biases under static scenarios, resulting in an environmental baseline vector. and computing power benchmark vector The data is then input into the built environment and run in a static matching function to perform a linear mapping operation. This function uses linear mapping and the calculation formula is as follows: The linear correlation between the environment and the baseline computing power is directly quantified through vector inner product operations, thereby outputting the environment-run static matching coefficient. If the environmental baseline vector and the computing power baseline vector are column vectors of the same dimension, then one of the vectors needs to be transposed (e.g., ...). This satisfies the dimension matching requirement of column vector × row vector in inner product operation, ensures the operation is valid, and outputs the environment-run static matching coefficient in scalar form.

[0035] Subsequently, the environment-operational static matching coefficient is subjected to second-order differential processing under a specified spatiotemporal dimension, and the cumulative change is solved using the second-order central difference method. The calculation formula is as follows: The cumulative changes in steady-state deviation of environmental parameters and offset from operating state reference are obtained. Finally, the accumulated changes are superimposed with the environmental baseline vector and the computing power baseline vector, according to the formula. To generate a steady-state feature vector that reflects the baseline synergistic relationship between the environment and operating conditions. This is denoted as the result of static coupling. Similarly, the transpose operation here is also to satisfy the horizontal format requirements of dimension stacking, ultimately generating a structurally regular steady-state feature vector.

[0036] Among them, the environmental baseline vector is used to reflect the steady-state distribution characteristics of the environmental perception parameters of the specified computing power facility in the same area and at the same time within the specified computing power facility; the computing power baseline vector is used to reflect the baseline state parameters of the corresponding operating equipment of the server cluster under the same computing power load pressure; the constructed environment-operation static matching function represents the objective function of maximizing the matching degree between the environment and the rated parameters of the equipment, which is used to quantify the static correspondence between the environmental perception parameters of the specified computing power facility and the standard operating state of the server cluster; the environment-operation static matching degree coefficient is used to quantify the degree of influence of the environmental perception parameters in the static environment of the specified computing power facility on the operating state of the server cluster.

[0037] In this embodiment, the static fusion process precisely adapts to the steady-state operation scenario of the computing power facility. Through min-max normalization and static correlation calibration, it effectively eliminates differences in data magnitude and system deviations, ensuring the reliability of the environmental baseline vector and the computing power baseline vector. The linear mapping operation based on vector inner product directly quantifies the linear correlation between the environment and the computing power baseline state without additional weighting, simplifying the calculation process while ensuring the accuracy of the matching coefficient. The cumulative change obtained through the second-order central difference method accurately captures the subtle cumulative effects of the environment and operating state under steady-state conditions. Combined with the steady-state feature vector generated by dimension superposition, it comprehensively integrates the baseline features and cumulative deviation information. This process complements the dynamic fusion process, fully covering both dynamic and static operation scenarios of the computing power facility, providing more comprehensive coupled data support for joint scheduling, and ensuring the efficient and stable operation of the server cluster in a steady-state environment.

[0038] Because the dynamic changes in environment and load during the operation of the computing cluster and the steady-state requirements of the baseline state need to be adapted separately, a quantitative judgment flowchart for joint scheduling, as shown in Figure 3, is provided based on the results of dynamic and static coupling processing. The quantitative judgment of joint scheduling is performed according to different scenarios: If it is a result of dynamic coupling processing, first determine whether the instantaneous rate of change exceeds the corresponding limit. If so, the CPU / GPU baseline frequency is increased to the upper limit of the interval and a load migration candidate path is generated; otherwise, the frequency adjustment gradient is maintained, and only the node load difference is monitored. If the difference exceeds the load balancing limit, cross-node load migration is performed. At the same time, the environment-operation coupling correlation coefficient is judged. If it is lower than the corresponding limit, migration is paused and the frequency is reduced. If it is a result of static coupling processing, first maintain the current CPU / GPU baseline frequency, and determine whether the node load difference exceeds the static limit. If so, resources are allocated according to priority and migration is performed; otherwise, the cumulative change is monitored. If it exceeds the limit, an imbalance warning is issued; if it does not exceed the limit, the operating parameters are maintained. This process achieves precise coordination of load, frequency, and migration through rapid response in dynamic scenarios and steady-state protection in static scenarios. It adapts to real-time load fluctuations and environmental changes while ensuring load balancing under baseline conditions, effectively improving the operational stability and resource utilization of the server cluster.

[0039] Further understanding is needed regarding the quantitative determination of joint scheduling, specifically including: if the coupling processing result is a dynamic coupling processing result, the output prioritizes triggering a rapid adjustment of the CPU / GPU base frequency to match load fluctuations, and simultaneously sends a cross-node load migration preparation command; if the instantaneous rate of change in the dynamic coupling processing result is greater than the preset instantaneous rate of change limit, based on the server hardware rated power and performance constraints, the CPU / GPU base frequency of the corresponding node is directly pulled to the upper limit of the preset load-adaptive frequency range to ensure rapid improvement of computing power to match the high dynamic change requirements; combined with the current load rate of each node and the heat distribution data between racks, screening... Nodes with loads lower than the current load rating (typically 50%-70% of server load, suitable for stable operation in most scenarios) are selected as migration targets. These nodes are sorted by communication latency from low to high to generate a list of cross-node load migration candidate paths, ensuring efficient and stable migration. Communication latency refers to the transmission time from the source node to the target node, measured in milliseconds (ms), based on the current load rate of each node and the heat distribution data between racks. This latency is obtained by synchronizing the node clocks and sending timestamped detection data packets, capturing the difference between the sending and receiving times of each node, and calculating the difference. The preset instantaneous change rate limit represents the maximum allowable response intensity dynamically correlated with the environment and operating status.

[0040] If the instantaneous rate of change in the dynamic coupling processing result is not greater than the preset instantaneous rate of change limit, the current CPU / GPU base frequency adjustment gradient is maintained, and only cross-node load migration condition monitoring and judgment are performed. At the same time, the node load difference between each server node in the server cluster is obtained. If the node load difference is greater than the preset load balancing limit, a formal cross-node load migration instruction is sent and executed based on the generated candidate path list. The preset load balancing limit represents the maximum allowable degree of load imbalance between server cluster nodes. If the node load difference is not greater than the preset load balancing limit, the real-time changes in the dynamic coupling processing result continue to be monitored.

[0041] The monitoring and judgment of cross-node load migration conditions are as follows: if the environment-run coupling correlation coefficient of the dynamic coupling processing result is lower than the preset environment-run coupling correlation coefficient limit, the cross-node load migration process is suspended and the CPU / GPU base frequency is lowered to the lower limit of the preset load adaptation frequency range. Otherwise, the current cross-node load migration state is maintained. The preset environment-run coupling correlation coefficient limit represents the minimum safety standard of the coordination relationship between the environment and the running state.

[0042] The preset parameters involved in the embodiments of the present invention include, but are not limited to, the upper limit of the preset load adaptation frequency range, the preset instantaneous change rate limit, the preset load balancing limit, the preset static load difference limit, the preset static cumulative change limit, and the lower limit of the preset load adaptation frequency range. All of these are pre-calibrated based on the rated parameters of the server cluster hardware, historical operating data, and the requirements of typical application scenarios. Furthermore, the specific values ​​of all preset parameters can be dynamically fine-tuned according to the actual scheduling scenario, ensuring the scientific nature, security, and adaptability of the system scheduling decision.

[0043] The quantitative determination for joint scheduling specifically includes: if the coupling processing result is a static coupling processing result, then the current CPU / GPU base frequency adjustment gradient is maintained; if the node load difference is greater than the preset static load difference limit, then the tasks to be migrated and the target nodes are sorted according to preset priorities (such as task urgency and node computing power redundancy), and then corresponding migration resources (bandwidth, computing power quota) are allocated. Task data and processes are migrated through cross-node data transmission protocols to achieve load balancing; the preset static load difference limit represents the maximum allowable deviation in load distribution between server cluster nodes; if the node load difference is not greater than the preset static load difference limit, then the cumulative change of the static coupling processing result continues to be monitored and determined. Specifically: if the cumulative change is greater than the preset static cumulative change limit, then the load between nodes is determined to be trending unbalanced, and an imbalance warning is issued; if the cumulative change is not greater than the preset static cumulative change limit, then the corresponding server cluster is prompted to maintain the current operating parameters unchanged; the preset static cumulative change limit represents the maximum allowable deviation in the relationship between the environment and the operating state baseline.

[0044] In this embodiment, by distinguishing between static coupling processing results and node load difference scenarios, precise adaptation of CPU / GPU frequency adjustment and cross-node load migration is achieved, ensuring stable operation while avoiding resource waste. Migration resources are allocated according to preset priorities, and standardized transmission protocols are used to ensure efficient and orderly load migration, quickly resolving load imbalance issues between nodes. Static cumulative change monitoring is introduced to proactively detect and warn of potential load imbalance trends, effectively maintaining load balance and operational stability among server cluster nodes, improving overall computing power utilization and task processing reliability, and adapting to cluster operation requirements in different scenarios.

[0045] Figure 4 shows the dynamic adjustment flowchart for power quotas of each server in the server cluster. The design logic is as follows: First, the collected power consumption data is normalized, and the median power consumption is obtained using the median sorting method. Then, the deviation value of each node is calculated. If the absolute value of the deviation exceeds a preset proportion, a first-level adjustment is triggered, adjusting the power quota in a step-by-step manner according to the excess range. After the first-level adjustment, if the power of a single node exceeds a preset proportion of the median power consumption, a second-level adjustment is triggered, increasing the quota adjustment range. For load migration scenarios, if the power of the target node exceeds a preset proportion of the median power consumption after migration, its quota is temporarily reduced. If the total power exceeds the rated value after the first-level adjustment, an exception notification is sent. This process, through hierarchical adjustment and dynamic constraints, achieves precise and balanced correction of the server cluster power quota, further ensuring balanced cluster power distribution and stable operation.

[0046] Further understanding is needed regarding the dynamic adjustment of power quotas for each server in the server cluster. Specifically, this involves: normalizing the currently collected power consumption data of the server cluster; calculating the median power consumption of the server cluster after normalization using a median sorting method; and calculating the deviation between the real-time power consumption of each server node and the median power consumption. If the absolute value of the deviation for any server node exceeds a preset percentage (usually set to ±15%), a first-level adjustment is triggered, and the power of the single node after the first-level adjustment is obtained. If the absolute value of the deviation for any server node does not exceed the preset percentage, then... Continue monitoring the operating status of each server node in the server cluster; Level 1 adjustment means: for server nodes whose absolute deviation value exceeds the preset proportion, adjust their power quota in a step-by-step manner according to the magnitude of the excess. For example, when the excess is 15%-25%, the power quota is reduced / increased by 8%; when it exceeds 25%-35%, the adjustment is increased to 15%; when it exceeds 35% or more, it is adjusted by 20%, and a 5% buffer space is reserved after each adjustment; if the power of a node is 30% higher than the median, the quota is directly increased by 20% to avoid excessive adjustment and fluctuations, and to ensure that it matches the actual load demand of the node.

[0047] If the power of a single node exceeds a preset percentage of the median power consumption, a secondary adjustment is triggered to coordinate the compatibility between the power quota and the primary adjustment strategy; if the power of a single node does not exceed a preset percentage of the median power consumption of the cluster, the load migration status is monitored and determined after the target node load migration.

[0048] Secondary adjustment means adding a preset adjustment range (usually set at 5%-10%) to the original tiered adjustment. For example, if the primary adjustment has already reduced by 8%, the secondary adjustment will add another 5% reduction, for a cumulative reduction of 13%; if the primary adjustment has increased by 15%, the secondary adjustment will add an 8% increase, for a cumulative increase of 23%; if a node is still 20% higher than the median after the primary adjustment, then an additional 10% will be added on top of the original 15% increase. By superimposing adjustments, the adaptability is strengthened, and the deviation from the median is quickly leveled off.

[0049] The monitoring and judgment of load migration status are as follows: If the real-time power consumption exceeds the preset percentage of the median power consumption, a temporary quota limit is imposed on the node, and the quota is reduced according to the preset percentage to avoid new power imbalance caused by load migration. For example, if it exceeds 15%-20%, the quota is reduced by 10%; if it exceeds 20%-25%, it is reduced by 15%; if it exceeds 25%, it is reduced by 20% and a load balancing warning is triggered. If the power of a node is 22% higher than the median, the quota is directly reduced by 15%, and new task access is restricted to avoid power overload after load migration and maintain the power balance of the cluster. If the real-time power consumption does not exceed the preset percentage of the median power consumption, the total power consumption of the server cluster after the first-level adjustment is monitored. If the total power consumption exceeds the rated total power, a power quota abnormality prompt is sent to prioritize the safe and stable operation of the server cluster. If the total power consumption does not exceed the rated total power (usually set to 85%-90% of the maximum carrying power of the server cluster, with redundancy space reserved to cope with sudden load), the power quota of each server in the server cluster is dynamically corrected.

[0050] In this embodiment, normalization and median deviation analysis are used to accurately locate power imbalance nodes. Combined with tiered primary adjustments and superimposed secondary adjustments, refined power quota adaptation is achieved, avoiding over- or under-adjustment. A temporary quota limitation mechanism effectively avoids new imbalances caused by load migration, while total power monitoring and anomaly alerts further ensure cluster security. The entire process relies on quantitative rules and tiered adjustment strategies to quickly level out node power deviations while maximizing adaptation to actual node load demands. This significantly improves server cluster power utilization efficiency and operational stability, extends hardware lifespan, and adapts to cluster operation requirements under high load fluctuation scenarios.

[0051] Furthermore, a gradual derating and collaborative derating determination is performed. Specifically, the operating data of the hierarchical scheduling scheme in the current scenario of the server cluster is compared with the corresponding reference operating data. The operating data includes the energy efficiency ratio and power fluctuation range, while the reference operating data includes the reference energy efficiency ratio and the power fluctuation range. The reference energy efficiency ratio is represented by the sum and average of historical energy efficiency ratios in the historical gradual derating determination process. If the obtained energy efficiency ratio is lower than the preset proportion of the reference energy efficiency ratio (e.g., 85%), or the obtained power fluctuation range exceeds the power fluctuation range (e.g., ±12%), then it is determined that gradual derating is required; otherwise, it is determined that collaborative derating is required. The fault tolerance rate is checked for the gradual and collaborative derating processes to determine whether the execution process of the hierarchical scheduling scheme needs to be suspended.

[0052] The progressive derating means: derating is reduced sequentially according to the hierarchical scheduling scheme, i.e.: the redundant resource scheduling layer is reduced by 20%-30% of its quota (e.g., closing 50% of idle memory channels, reducing redundant fan speeds to the minimum operating threshold), the edge load balancing scheduling layer is reduced by 15%-20% (e.g., limiting the number of concurrent tasks per node to 80% of the original quota, reducing the frequency of cross-node data synchronization), and the task scheduling layer only reduces the quota for non-core tasks by 5%-10% (e.g., the computational resource ratio of compressed log analysis and backup tasks). Simultaneously, within a preset derating period (e.g., 5 minutes), the number of task failures at each layer is counted. The collaborative derating means: based on the total power target of the server cluster (e.g., 80% of the rated power), the derating of each scheduling layer in the hierarchical scheduling scheme is reduced proportionally (e.g., a uniform reduction of 10%). Simultaneously, within a preset derating period, the number of cross-layer data interaction failures is counted.

[0053] The fault tolerance rate verification is as follows: During the gradual derating process, if the number of task failures in any scheduling layer exceeds the allowed number of failures for that scheduling layer within a preset derating period (5% of the total number of tasks in that layer, i.e., a single-layer fault tolerance rate of less than 95%), the execution process of that scheduling layer must be paused, and the fault tolerance rate verification must be performed again after sending a power quota callback prompt. During the collaborative derating process, if the number of cross-layer data interaction failures within a preset derating period exceeds the allowed number of interaction failures (usually set to 2%, i.e., a cross-layer fault tolerance rate of less than 98%), the execution process of the hierarchical scheduling scheme must be paused, and the fault tolerance rate verification must be performed again after sending a prompt to adjust the quota allocation ratio of each scheduling layer. If no pause operation is performed during the above derating processes, the current hierarchical scheduling scheme is deemed to be suitable for the current scheduling scenario, and the corresponding execution process continues to be monitored.

[0054] The specific values ​​in the above steps (such as the preset derating period, single-level fault tolerance rate, cross-level fault tolerance rate, and quota adjustment ratio) are usually set based on the server cluster hardware performance parameters, historical operating data, and load scenario characteristics. For example, the specific values ​​of 5% and 2% involved in the fault tolerance rate verification process are determined by preset personnel based on the historical failure probability distribution of the server cluster under historical computing power load, the tolerance of historical task execution for errors, and hardware redundancy capabilities. 5% adapts to single-level local fault tolerance requirements, while 2% meets the high stability requirements of cross-level collaboration. In actual deployment and operation scenarios, these values ​​can be flexibly fine-tuned according to real-time load fluctuations, task priority requirements, and cluster stability goals to ensure that the derating judgment and fault tolerance verification are more in line with actual operating needs and improve the adaptability of the solution.

[0055] In this embodiment, the derating strategy is precisely adapted by using a dual-dimensional approach of energy efficiency ratio and power fluctuation based on historical data. Gradual, layered derating prioritizes compressing non-core resources, while coordinated derating focuses on the total power target, ensuring the stability of core tasks while avoiding resource waste. A dual verification mechanism of layered and cross-layer fault tolerance is introduced to effectively manage the risks of task failures and abnormal data interaction. Actions such as pausing processes and reverting quotas are used to promptly prevent the spread of faults, improving the robustness of the solution. The entire process, based on data comparison, using tiered derating as a means, and fault tolerance verification as a guarantee, achieves a balance between the scientific nature and security of the derating strategy, providing strong support for the high-load, high-reliability operation of the server cluster.

[0056] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0057] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0058] In various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0059] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0060] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic scheduling system for computing resources that integrates environmental perception and power self-balancing, characterized in that, The system includes: a data acquisition and environmental perception fusion module, used to collect multi-dimensional data from inside and outside the designated computing facilities running the server cluster in real time through a sensor network, and to fuse it with current environmental perception parameters; a joint scheduling and self-power balancing control module, used to make quantitative judgments on joint scheduling based on the fusion processing results, and to dynamically correct the power quotas of each server in the server cluster based on the acquired power consumption data; and a hierarchical scheduling and multi-scenario adaptation verification module, used to generate a hierarchical scheduling scheme based on the power quota constraints, and to make judgments on gradual and collaborative derating, while verifying the fault tolerance rate of the derating process. The hierarchical scheduling scheme includes a redundant resource scheduling layer and an edge load balancing layer. The system comprises a balance scheduling layer and a task scheduling layer. The specific process of the fusion processing includes: classifying the collected multi-dimensional data into data types under different spatiotemporal dimensions to obtain environmental dimension data and computing power dimension data; performing standardization and spatiotemporal alignment processing to obtain environmental feature vectors and computing power feature vectors; inputting these vectors into a pre-constructed environment-running coupling function for nonlinear mapping to output the environment-running coupling correlation coefficient; performing first-order differential processing on the environment-running coupling correlation coefficient in continuous spatiotemporal dimensions to obtain the instantaneous change rate of environmental parameter changes and operational state response; and concatenating the obtained instantaneous change rate with the environmental feature vector and computing power feature vector to generate a data structure reflecting the environment-running coupling correlation coefficient. The real-time feature vector of the collaborative relationship between the environment and the operating state is denoted as the dynamic coupling processing result; the environment feature vector is used to reflect the dynamic distribution of the environment perception parameters of the specified computing power facility under different regions and different operating periods within the specified computing power facility; the computing power feature vector is used to reflect the real-time changes of the corresponding operating devices of the server cluster under different computing power load pressures; the constructed environment-operation coupling function is used to quantify the dynamic correlation between the environment perception parameters of the specified computing power facility and the operating state of the server cluster; the environment-operation coupling correlation coefficient is used to quantify the degree of influence of the environment perception parameters in the dynamic environment of the specified computing power facility on the operating state of the server cluster; the coupling... The combined processing also includes: normalizing and statically correlating the collected multi-dimensional data under the same spatiotemporal dimension to obtain the environmental reference vector and the computing power reference vector, and inputting them into the constructed environment-operation static matching function for linear mapping to output the environment-operation static matching degree coefficient; performing second-order differential processing on the environment-operation static matching degree coefficient under the specified spatiotemporal dimension to obtain the cumulative change of the steady-state deviation of environmental parameters and the benchmark offset of operating state; and performing dimensional superposition processing on the obtained cumulative change with the environmental reference vector and the computing power reference vector to generate a steady-state feature vector that reflects the benchmark coordination relationship between the environment and the operating state, which is recorded as the static coupling processing result.

2. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 1, characterized in that, The environmental reference vector is used to reflect the steady-state distribution characteristics of the environmental perception parameters of the specified computing facility in the same area and at the same time period within the specified computing facility. The computing power benchmark vector is used to reflect the benchmark state parameters of the corresponding operating equipment of the server cluster under the same computing power load pressure; the constructed environment-operation static matching function represents the objective function of maximizing the matching degree between the environment and the rated parameters of the equipment, and is used to quantify the static correspondence between the environmental perception parameters of the specified computing power facility and the standard operating state of the server cluster; the environment-operation static matching degree coefficient is used to quantify the degree of influence of the environmental perception parameters of the static environment in which the specified computing power facility is located on the operating state of the server cluster.

3. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 1, characterized in that, The joint scheduling and self-power balancing control module includes a joint scheduling and judgment unit and a self-power balancing control unit. The joint scheduling and judgment unit receives the coupling processing results and, in conjunction with the current load distribution and task priority requirements of the server cluster, performs quantitative judgment on joint scheduling to output joint scheduling instructions that meet performance constraints. The coupling processing results include dynamic coupling processing results and static coupling processing results. The joint scheduling includes dynamic adjustment of the CPU / GPU base frequency to match real-time load fluctuations and adapt to dynamic environmental changes, and cross-node load migration to balance node load pressure and avoid local overheating. The self-power balancing control unit, based on the output joint scheduling instructions and real-time power consumption data, dynamically corrects the power quota of each server in the server cluster under the closed-loop feedback control logic of the server cluster, ensuring that the server cluster maintains a dynamic balance between performance output and power consumption. The power consumption data reflects the dynamic changes in the real-time operating power consumption of each server in the server cluster, the peak power of a single node, and the total energy consumption of the entire cluster.

4. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 3, characterized in that, The quantitative determination for joint scheduling specifically includes: if the coupling processing result is a dynamic coupling processing result, then the CPU / GPU base frequency is preferentially triggered to quickly adjust to match load fluctuations, and a cross-node load migration preparation command is sent simultaneously; if the instantaneous change rate in the dynamic coupling processing result is greater than a preset instantaneous change rate limit, then the CPU / GPU base frequency of the corresponding node is increased to the upper limit of the preset load adaptation frequency range, and a candidate path list for cross-node load migration is generated, where the preset instantaneous change rate limit represents the maximum allowable response intensity dynamically associated with the environment and operating status; if the instantaneous change rate in the dynamic coupling processing result is not greater than the preset instantaneous change rate limit, then the current CPU / GPU base frequency adjustment gradient is maintained, and only cross-node load migration condition monitoring and determination are performed, while obtaining the node information between each server node in the server cluster. Load difference; if the node load difference is greater than the preset load balancing limit, a formal cross-node load migration instruction is sent and executed based on the generated candidate path list. The preset load balancing limit represents the maximum allowable imbalance in load distribution among server cluster nodes. If the node load difference is not greater than the preset load balancing limit, the real-time changes in the dynamic coupling processing result continue to be monitored. The cross-node load migration condition monitoring and judgment are specifically as follows: if the environment-running coupling correlation coefficient of the dynamic coupling processing result is lower than the preset environment-running coupling correlation coefficient limit, the cross-node load migration process is paused, and the CPU / GPU base frequency is lowered to the lower limit of the preset load adaptation frequency range. Otherwise, the current cross-node load migration state is maintained. The preset environment-running coupling correlation coefficient limit represents the minimum safety standard for the collaborative relationship between the environment and the running state.

5. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 4, characterized in that, The quantitative determination for joint scheduling further includes: if the coupling processing result is a static coupling processing result, then maintaining the current CPU / GPU base frequency adjustment gradient; if the node load difference is greater than a preset static load difference limit, then allocating migration resources according to a preset priority order and performing cross-node load migration, where the preset static load difference limit represents the maximum allowable deviation in load distribution between server cluster nodes; if the node load difference is not greater than the preset static load difference limit, then continuing to monitor and determine the cumulative change in the static coupling processing result, specifically: if the cumulative change is greater than a preset static cumulative change limit, then determining that the load between nodes is trending towards imbalance and issuing an imbalance warning; if the cumulative change is not greater than a preset static cumulative change limit, then prompting the corresponding server cluster to maintain its current operating parameters unchanged; where the preset static cumulative change limit represents the maximum allowable deviation in the relationship between the environment and the operating state baseline.

6. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 3, characterized in that, The dynamic correction of power quotas for each server in the server cluster specifically includes: normalizing the power consumption data currently collected by the server cluster, calculating the median power consumption of the server cluster after normalization using the median sorting method, and calculating the deviation between the real-time power consumption of each server node in the server cluster and the median power consumption; if the absolute value of the deviation of a server node exceeds a preset proportion, a first-level adjustment is triggered and the power of the single node after the first-level adjustment is obtained; if the absolute value of the deviation of no server node exceeds the preset proportion, the operating status of each server node in the server cluster continues to be monitored; the first-level adjustment means that for server nodes whose absolute value of deviation exceeds the preset proportion, their power quota is adjusted in steps according to the magnitude of the excess.

7. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 6, characterized in that, The dynamic correction of power quotas for each server in the server cluster further includes: if the power of a single node exceeds a preset percentage of the median power consumption, a secondary adjustment is triggered to coordinate the compatibility between power quotas and the primary adjustment strategy; if the power of a single node does not exceed a preset percentage of the median power consumption of the cluster, the load migration status is monitored and determined after the target node is migrated; the secondary adjustment means adding a preset percentage of quota adjustment based on the original tiered adjustment; the monitoring and determination of the load migration status specifically involves: if the real-time power consumption exceeds a preset percentage of the median power consumption, a temporary quota limit is imposed on the node, and the quota is reduced by the preset percentage to avoid new power imbalances caused by load migration; if the real-time power consumption does not exceed a preset percentage of the median power consumption, the total power consumption of the server cluster after the primary adjustment is monitored; if the total power consumption exceeds the rated total power, a power quota anomaly warning is sent to prioritize the safe and stable operation of the server cluster; if the total power consumption does not exceed the rated total power, the dynamic correction of power quotas for each server in the server cluster is completed.

8. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 7, characterized in that, The redundant resource scheduling layer, edge load balancing scheduling layer, and task scheduling layer allocate power quotas and perform computing load processes in a progressive manner. The progressive derating and collaborative derating determination is specifically as follows: the operating data of the hierarchical scheduling scheme in the current scenario of the server cluster is compared with the corresponding reference operating data. The operating data includes the energy efficiency ratio and power fluctuation range, and the reference operating data includes the reference energy efficiency ratio and the power fluctuation range. If the obtained energy efficiency ratio is lower than the preset ratio of the reference energy efficiency ratio, or the obtained power fluctuation range exceeds the power fluctuation range, then a progressive derating determination is made. Conversely, if the conditions are not met, a coordinated reduction in spending will be determined. The fault tolerance rate is checked for the gradual and collaborative derating process to determine whether to suspend the execution flow of the hierarchical scheduling scheme; the gradual derating means: derating is reduced in sequence according to the hierarchical scheduling scheme, while counting the number of task failures at each layer within a preset derating period. The collaborative derating means that, based on the total power target of the server cluster, the derating of each scheduling layer in the hierarchical scheduling scheme is reduced proportionally, and the number of cross-layer data interaction failures is counted within a preset derating period.

9. The dynamic scheduling system for computing resources integrating environmental perception and power self-balancing as described in claim 8, characterized in that, The fault tolerance rate verification is specifically as follows: during the gradual derating process, if the number of task failures in any scheduling layer exceeds the allowed number of failures for that scheduling layer within the preset derating period, the execution process corresponding to that scheduling layer is suspended, and the fault tolerance rate verification is performed again after sending the power quota callback prompt. During the collaborative reduction process, if the number of cross-layer data interaction failures within the preset reduction period exceeds the allowed number of interaction failures, the execution process of the hierarchical scheduling scheme will be suspended, and the fault tolerance rate will be re-verified after sending a prompt to adjust the quota allocation ratio of each scheduling layer. If the above reduction process does not involve suspending the execution process, it is determined that the current hierarchical scheduling scheme is suitable for the current scheduling scenario, and the corresponding execution process will continue to be monitored.

Images