A marginal energy efficiency scheduling system for computing power server joint game

Abstract

The present application belongs to the technical field of computing power energy efficiency scheduling, and provides a marginal energy efficiency scheduling system for computing power server joint game, comprising: mining cluster historical load characteristics, identifying multi-tenant and queue resource competition conflict, and constructing a dynamic load game field with a competition penalty mechanism; building an energy efficiency-delay multi-dimensional time sequence monitoring system, which can identify performance false prosperity anomalies that energy efficiency is lifted but response delay approaches the service level agreement bottom line; relying on the false prosperity state to carry out load intensity climbing test, positioning the performance collapse critical point and the energy efficiency yield decreasing inflection point, and delimiting the optimization effective scope considering performance and energy efficiency; distinguishing two types of invalid scheduling strategies, response lag and energy efficiency overdraft, and respectively adjusting the monitoring alarm threshold and the game field energy consumption cost coefficient to complete adaptive correction, so that the present application solves the energy efficiency and performance imbalance problem under multi-subject resource competition, and realizes the closed-loop management and control of the whole process of energy efficiency scheduling.

Description

Technical Field

[0001] This invention belongs to the field of computing power energy efficiency scheduling technology, specifically a marginal energy efficiency scheduling system based on joint game theory of computing power servers. Background Technology

[0002] As the scale of computing clusters continues to expand, resource competition among multiple tenants and service queues is becoming increasingly fierce. How to improve overall energy efficiency while ensuring service quality has become a key issue in the field of computing scheduling.

[0003] However, existing energy efficiency scheduling strategies for computing servers generally suffer from the following prominent problems: First, there is a lack of unified modeling for load characteristics and multi-entity resource competition relationships, and energy efficiency optimization and performance assurance are often treated separately, making it difficult to support collaborative scheduling decisions in joint game scenarios. Second, energy efficiency assessment methods mostly rely on steady-state index analysis, failing to effectively capture the false boom state of energy efficiency temporarily rising while performance simultaneously deteriorates under dynamic loads, and lacking real-time monitoring and anomaly identification capabilities for the co-evolution of energy efficiency and performance. Third, energy efficiency scheduling decisions lack a systematic method for defining safety boundaries, i.e., it is impossible to clearly define to what extent energy efficiency optimization will trigger performance collapse, causing the system to easily exceed the critical point without realizing it when pursuing energy efficiency improvements. Fourth, when the scheduling strategy deviates from the safety boundary, existing methods lack classification and identification mechanisms for failure tendencies and differentiated correction mechanisms, making it difficult to achieve closed-loop governance from anomaly detection to strategy adjustment. The existence of these technical problems makes computing clusters face significant risks of efficiency and performance imbalance in joint game scenarios.

[0004] Therefore, this invention provides a marginal energy efficiency scheduling system for joint game theory among computing servers. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a marginal energy efficiency scheduling system for joint game theory of computing power servers, comprising the following modules: Load game construction module: By tracking the historical task load sequence of the computing power cluster, extracting load characteristics and identifying the resource demand distribution and conflict relationship of multiple subjects, a dynamic load game field containing load characteristics and subject competition relationship is constructed. Performance and latency monitoring module: Under the operating conditions of dynamic load game field, collect the real-time energy efficiency ratio and task response latency of each node, and establish a multi-dimensional monitoring system for energy efficiency and latency: when the task response latency approaches the service level agreement bottom line while the overall system energy efficiency still shows an upward trend, it is judged as a false prosperity state of performance. Optimize the safety definition module: Based on the false prosperity state of efficiency, implement a dynamic load intensity increase strategy, track the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits, and define the effective scope of energy efficiency optimization of energy efficiency scheduling strategy without triggering performance degradation. The optimization strategy correction module identifies the failure tendency of existing scheduling strategies that deviate from the effective scope of energy efficiency optimization, and classifies and optimizes existing scheduling strategies: for strategies with delayed response, the multi-dimensional monitoring system is adjusted to detect the critical risk of approaching performance collapse earlier; for strategies with energy efficiency overdraft, the dynamic load game field is adjusted so that scheduling decisions are proactively withdrawn before the inflection point of diminishing energy efficiency benefits is crossed.

[0007] As a further aspect of the present invention: the method of extracting load features and identifying the distribution and conflict relationships of resource demands among multiple entities by tracking the historical task load sequences of the computing power cluster is as follows: Export historical task execution records from the scheduling system logs, retaining records of tasks that were completed normally; Calculate task waiting time, runtime, and resource request volume; group by tenant or queue identifier to count the total number of tasks, average and maximum resource request volume, and task arrival time sequence for each subject; divide into hourly segments to count the total resource request volume for each subject; if the total resource request volume exceeds the total capacity of the same type of resources in the cluster, it is determined that there is a competition conflict and the conflicting subject pair and conflict time period are recorded.

[0008] As a further aspect of the present invention, the method for constructing a dynamic load game field is as follows: Each tenant or queue is defined as a participating entity, which has independent decision-making power for resource requests. The system status is described by the current availability of three types of resources: CPU, memory, and GPU. The current availability is equal to the total resource capacity minus the occupied amount. Each entity submits a resource request based on the current system status in each scheduling time slot. The scheduler aggregates the requests and allocates resources according to the requests. After allocation, the system status is updated. The system status is restored after the task is completed and resources are released.

[0009] As a further aspect of the present invention: in a dynamic load game, the utility function of each subject consists of a benefit term and a cost term; the benefit term is calculated by multiplying the square of the ratio of the actual allocated resources to the requested resources by a benchmark value, and the larger the allocation gap, the faster the loss of benefits. The cost item includes the sum of completion time cost and energy consumption cost. The completion time cost equals the task runtime multiplied by the time cost coefficient, and the energy consumption cost equals the total energy consumption of the cluster multiplied by the energy consumption cost coefficient and is shared equally by all entities. The utility value equals the benefit item minus the completion time cost and then minus the energy consumption cost.

[0010] As a further aspect of the present invention: A multi-party competition penalty rule is set in the dynamic load game field: Using scheduling time slots as the unit window, if multiple entities with conflicting relationships apply for resources in any type of resource and their application amounts all exceed the set percentage of their respective baseline values, a competition penalty will be triggered. After triggering, the energy consumption cost of each triggering entity is increased by an additional penalty value based on the current value. The additional penalty value is equal to the current energy consumption cost multiplied by the penalty coefficient. Each triggering entity calculates and adds the penalty value to its own energy consumption cost independently.

[0011] As a further aspect of the present invention: the method for collecting the real-time energy efficiency ratio and task response latency of each node is as follows: Using each computing server as the data collection object, the instantaneous power consumption of the node is read through the hardware management interface, and the computing power output value of the node is read through the operating system interface. The real-time energy efficiency ratio is equal to the computing power output value divided by the power consumption value. The submission time of all running tasks on the node is obtained from the task queue of the scheduling system, and the response latency is calculated. The average value of the response latency of all running tasks on the node is taken as the representative value of the node's current collection period.

[0012] As a further aspect of the present invention, the method for establishing a multi-dimensional monitoring system is as follows: The monitoring system includes energy efficiency and latency dimensions. For each node, two time series, energy efficiency and latency, are maintained and updated on a rolling basis. The energy efficiency ratio and response latency values ​​obtained by the same node in the same acquisition cycle are marked as paired data points under the same timestamp. The pairing relationship ensures that changes in energy efficiency and latency can be correlated and compared on the same time base in subsequent analysis.

[0013] As a further aspect of the present invention: the method for determining a false prosperity state of performance is as follows: After each data collection cycle is completed, the average response delay of all nodes is calculated. If the average response delay reaches the set percentage of the service level agreement's minimum value, the approximation condition is triggered. At the same time, the average system energy efficiency of the current and the previous five data collection cycles is taken to determine the trend. If the slope of the fitted line is greater than zero, the energy efficiency is determined to be on an upward trend. When the approximation condition and the upward trend of energy efficiency are met simultaneously in the same data collection cycle, it is determined to be a false prosperity state of efficiency.

[0014] As a further aspect of the present invention: the method for tracking the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits is as follows: Starting with the load intensity at the point of triggering a false boom state, the load intensity is increased in increments of a percentage from the initial value. After each round of stable operation, the average response latency and average energy efficiency ratio are collected. The load intensity of the previous round before the average response latency first reaches or exceeds the service level agreement threshold is recorded as the performance collapse critical point. The load intensity of the previous round before the average energy efficiency ratio first falls below the previous round is recorded as the energy efficiency benefit reduction inflection point. The smaller of the two corresponding load intensities is taken as the upper limit of the effective scope, and the initial load intensity is taken as the lower limit of the effective scope.

[0015] As a further aspect of the present invention, the process of dividing and optimizing the existing scheduling strategy is as follows: Compare the current actual load intensity with the effective scope. If the actual load intensity is less than the lower limit of the scope, it is determined to be a response lag strategy; if it is greater than the upper limit of the scope, it is determined to be an energy efficiency overdraft strategy. For response lag strategies, lower the monitoring and alarm threshold so that the alarm trigger value is close to the actual critical value. For energy efficiency overdraft strategies, increase the energy consumption cost coefficient in the dynamic load game field until the load intensity corresponding to the system pullback point is equal to or less than the load intensity corresponding to the energy efficiency reduction inflection point.

[0016] The beneficial effects of this invention are as follows: The load game construction module uniformly models load characteristics and multi-subject resource conflicts, builds a dynamic game field to quantify the decision-making benefits and competitive penalties of subject resources, integrates resource competition relationships into the scheduling layer, realizes energy efficiency collaborative decision-making in competitive scenarios, and eliminates the shortcomings of separate scheduling of the two; The efficiency and latency monitoring module establishes a synchronous time-series monitoring mechanism for energy efficiency and latency, and judges deviations such as latency approaching the service level agreement bottom line and energy efficiency continuously rising, providing real-time early warning of the performance degradation risk brought about by energy efficiency optimization, and making up for the lag in steady-state index evaluation; The optimization safety definition module quantitatively obtains the performance collapse and energy efficiency benefit inflection point through gradient load intensity tests, delineates the safe load range, provides a quantitative safety basis for scheduling decisions, and avoids system degradation caused by blindly pursuing energy efficiency; The optimization strategy correction module distinguishes two types of failure scheduling strategies, respectively lowering the monitoring alarm threshold to detect critical risks in advance and raising the game energy consumption cost coefficient to prompt the load to be withdrawn in advance, forming a closed-loop adaptive scheduling of anomaly identification-classification correction, and continuously balancing system energy efficiency and service quality. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a marginal energy efficiency scheduling system for joint game theory of computing power servers according to the present invention. Figure 2 This is a schematic diagram of the distribution of main resource demand and resource conflict relationship in this invention; Figure 3 This is a schematic diagram of the present invention regarding the monitoring and determination of false prosperity of energy efficiency ratio and response delay time series; Figure 4 This is a schematic diagram of the performance collapse critical point and the inflection point of diminishing energy efficiency benefits of the load intensity dynamic climbing tracking system of this invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] Example: Please refer to Figures 1-4 As shown in the embodiment of the present invention, a marginal energy efficiency scheduling system for joint game theory of computing power servers includes the following modules: Load game construction module: By tracking the historical task load sequence of the computing power cluster, extracting load characteristics and identifying the resource demand distribution and conflict relationship of multiple subjects, a dynamic load game field containing load characteristics and subject competition relationship is constructed. In the load game construction module, the process of extracting load characteristics and identifying the resource demand distribution and conflict relationships of multiple stakeholders by tracking the historical task load sequence of the computing power cluster is as follows: Export historical task execution records from the scheduling system logs of the computing power cluster. Each record should contain at least the following fields: task submission time, task start execution time, task end time, number of CPU cores requested, number of memory requested, number of GPUs requested, and tenant name or service queue identifier to which the task belongs. Remove task records marked as abnormal interruption or manually terminated by the user, and retain task records marked as normal completion; Extract the load characteristic information of each historical task. For each task record retained after cleaning, calculate the following information in sequence: Task waiting time: equal to the task start time minus the task submission time; Task runtime: equal to the task end time minus the task start time; Task resource request includes three dimensions: CPU cores, memory capacity, and GPUs. Associate the information of each task with the tenant or service queue identifier to which the task belongs and store it. Resource demand distribution is aggregated by subject dimension. All task records are grouped according to tenant or service queue identifier. All tasks of the same tenant or queue are grouped into the same subject group. For each subject group, the following information is collected: total number of tasks submitted by the subject; average and maximum CPU core requests for all tasks in the subject; average and maximum memory capacity requests for all tasks in the subject; average and maximum GPU requests for all tasks in the subject; average and maximum runtime for all tasks in the subject; and the time series of task arrivals, i.e., the number of tasks submitted in each hour of the day, as exemplified in Table 1 below. Table 1: Distribution of Resource Demand; ; To identify resource demand conflicts among multiple entities, a day is divided into 24 time periods, each lasting one hour. For each time period, the number of tasks submitted by all entities and the total amount of various resources requested by the tasks are counted. For CPU resources, if the total number of CPU cores requested by multiple entities in the same time period exceeds the total number of CPU cores in the cluster, it is determined that multiple entities have a competition conflict for CPU resources in the time period. The same determination method is used for memory resources and GPU resources. Each pair of conflicting entities, the time period of conflict, and the type of conflicting resources are recorded, as exemplified in Table 2 below. Table 2: Resource Demand Conflicts Among Entities; ; In the load game construction module, the process of constructing a dynamic load game field that includes load characteristics and the competitive relationships between the subjects is as follows: Define the participating entities in the game arena. Each tenant or each service queue is defined as a participating entity. The number of entities is equal to the sum of the number of tenants and queues in the cluster. Each entity has independent resource request decision-making power in the game arena. That is, the entity can decide independently the number of CPU cores, memory capacity and GPUs requested in each round of scheduling according to its own needs. Define the system state space. The system state is described by the current availability of various resources in the cluster. The resource types include three categories: CPU cores, memory, and GPU. At any point in time, the system state is calculated as follows: the system state equals the total capacity of various resources in the cluster minus the amount of resources occupied by running tasks. For example, if the cluster has 100 CPU cores and currently running tasks have occupied 40 cores, then the current available CPU cores are 60. Define a utility function for each agent. The utility function quantifies the overall gain each agent receives in the game. Each agent's goal is to maximize its own utility function value. The utility function consists of two parts: a gain term and a cost term. The benefit item reflects the value gained by an entity after acquiring resources. For example, if an entity requests X CPU cores in a scheduling round and is actually allocated Y CPU cores, and Y equals X, meaning the demand is fully met, the benefit item is 100 points. If Y is less than X, the benefit item is calculated by multiplying the square of (Y divided by X) by 100 points. For example, if an entity requests 10 cores and is actually allocated 8 cores, the benefit item is equal to the square of (8 divided by 10) multiplied by 100, which equals 64 points. This ensures that the larger the allocation gap, the faster the benefit is lost, thus prompting the scheduler to try to meet the entity's demand. The cost item consists of two parts. The first part is the completion time cost, which is equal to the task runtime multiplied by the time cost coefficient. For example, if the time cost coefficient is set to 1 point per second, the completion time cost is 10 points if the task runs for 10 seconds. The second part is the energy consumption cost, which is equal to the total energy consumption of the cluster in this round of scheduling multiplied by the energy consumption cost coefficient, and is shared equally among all entities. For example, if the total energy consumption of the cluster in this round is 50 kWh and the energy consumption cost coefficient is 0.3 points per kWh, then the energy consumption cost is 15 points, and each entity bears 15 points. The final calculation method for utility value is as follows: utility value equals the benefit item minus the completion time cost minus the energy consumption cost. For example: if a subject requests 10 CPU cores and is actually allocated 10 cores, the benefit item is 100 points; the task runs for 10 seconds, the completion time cost is 10 points; the total energy consumption of this round of cluster is 50 kWh, the energy consumption cost is 15 points; the subject's utility value equals 100 minus 10 minus 15, which equals 75 points; if the same subject is actually allocated only 8 cores, the benefit item drops to 64 points, and other conditions remain unchanged, then the utility value equals 64 minus 10 minus 15, which equals 39 points. Establish a coupling relationship between the subject strategy and the system state. At the beginning of each scheduling time slot, each subject submits its own resource request based on the current system state. After the requests of all subjects are aggregated, the scheduler allocates resources according to the requests of each subject. After allocation, the system state is updated accordingly, and the current available amount of various resources is reduced accordingly. When the allocated tasks complete their execution and release resources, the system state is restored accordingly. Establish dynamic operating rules for the game arena. In the continuous operation of the game arena, the process of application-allocation-release is executed cyclically. Each subject's strategy choice is affected by two factors: first, the subject's own resource demand distribution characteristics, which are determined by historical statistical data extracted from previous steps; second, the resource application behavior of other subjects. The more other subjects apply, the less remaining resources the current subject has, which in turn affects the current subject's utility value. The mutual constraint relationship between subjects is the game-like expression of the subject competition relationship. Load characteristics and conflict relationships are loaded into the game field. The resource demand distribution statistics of each subject extracted in the previous step serve as a reference baseline for the amount of resources requested by the subject in the game. The conflict relationships between subjects identified in the previous step are used to set the multi-subject competition penalty rules. The triggering condition for the competition penalty rule is as follows: a unit window time length is set, which is equal to a scheduling time slot, usually 1 minute. Within a unit window time, for any type of resource, if multiple subjects with conflicting relationships each request more than 80% of their respective baseline values ​​for that resource, then the competition penalty is triggered. For example, if subject A and subject B have historically requested 10 and 15 CPU cores respectively, and within the window, subject A requests 9 cores and subject B requests 13 cores, then 9 exceeds 80% of 10 (i.e., 8 cores), and 13 exceeds 80% of 15 (i.e., 12 cores). If both conditions are met simultaneously, then the competition penalty is triggered. The competition penalty is calculated as follows: After the competition penalty is triggered, each triggering entity's energy consumption cost item is increased by an additional penalty value based on the current value. The additional penalty value is equal to the triggering entity's current energy consumption cost multiplied by a penalty coefficient. The penalty coefficient ranges from 0 to 0.5, and the specific value is preset by the cluster administrator. For example, if the penalty coefficient is set to 0.3 and the triggering entity's current energy consumption cost is 15 points, then the entity's additional penalty value is 15 multiplied by 0.3, which equals 4.5 points. Its final energy consumption cost becomes 15 plus 4.5, which equals 19.5 points. Each entity that triggers the competition penalty calculates its own additional penalty value independently and adds it to its own energy consumption cost, rather than all entities sharing a total amount. Finally, the construction of the dynamic load game field is completed, as shown in Table 3 below for example; Table 3: Elements of Dynamic Load Game Field; ; It is understandable that the functional significance of the load game construction module is: by integrating historical load characteristics with the resource competition relationship of multiple subjects, a dynamic game field is constructed, which transforms the traditional passive resource allocation into a competitive environment in which multiple subjects make autonomous decisions. This provides an evolvable basic simulation platform for subsequent energy efficiency scheduling analysis, so that energy efficiency optimization is no longer isolated from the resource competition relationship, and solves the defect of the separation between energy efficiency and performance from the source. Performance and latency monitoring module: Under dynamic load game field operation conditions, collect the real-time energy efficiency ratio and task response latency of each node to establish a multi-dimensional monitoring system for energy efficiency and latency; when the task response latency approaches the service level agreement bottom line while the overall system energy efficiency still shows an upward trend, it is judged as a false prosperity state of performance. In the performance latency monitoring module, under the operating conditions of a dynamic load game, the process of collecting the real-time energy efficiency ratio and task response latency of each node is as follows: Each physical node in the cluster (i.e., each computing server) is treated as a data collection object; the collection cycle is set to be executed once every 10 seconds, and each collection covers all nodes in the cluster; During each acquisition cycle, the instantaneous power consumption of the node at the current moment is read through the node's hardware management interface, in watts; the computing power output value of the node at the current moment is read through the node's operating system interface, in floating-point operations per second. The real-time energy efficiency ratio of the node is equal to the computing power output value divided by the power consumption value. Within each collection cycle, all running task records on the node are retrieved from the task queue of the scheduling system. For each running task, the submission time and the current time are extracted. The task response delay is equal to the current time minus the submission time. The average response delay of all running tasks on the node is taken as the representative value of the node's task response delay in the current collection cycle. In the performance delay monitoring module, the process of establishing a multi-dimensional monitoring system for energy efficiency and delay is as follows: The multi-dimensional monitoring system includes two core dimensions: energy efficiency and latency. The energy efficiency dimension records the real-time energy efficiency ratio of each node in each collection cycle, which reflects the energy utilization efficiency of the node. The latency dimension records the task response latency of each node in each collection cycle, which reflects the service quality level of the node. For each node, two time series are maintained: energy efficiency and latency. After each data collection, new data points are appended to the end of the corresponding dimension of the time series of the corresponding node, and old data points that exceed the set retention period (e.g., 24 hours) are automatically discarded. For each acquisition time, the energy efficiency ratio and response delay values ​​acquired by the same node within the same acquisition cycle are marked as paired data points under the same timestamp. The pairing relationship ensures that the changes in energy efficiency and time delay in subsequent analysis can be correlated and compared on the same time base. In the performance latency monitoring module, the specific process for determining a state of false performance prosperity is as follows: The service level agreement baseline is preset by the cluster administrator according to business needs. For example, setting the average response latency of all nodes to not exceed 200 milliseconds is the baseline. After each data collection cycle is completed, the average response latency of all node tasks at the current time is calculated. If the average response latency of the tasks is greater than or equal to 90% of the baseline value, the approximation condition is triggered. The average system energy efficiency is taken as the current acquisition cycle and the previous five consecutive acquisition cycles (i.e., a total of 6 data points in the past 60 seconds). The average system energy efficiency is equal to the sum of the energy efficiency ratios of all nodes divided by the total number of nodes. The least squares method is used to perform linear fitting on the average system energy efficiency of the acquisition cycle. If the slope of the fitted line is greater than 0, it is determined that the overall system energy efficiency is on the rise. If the approach condition is triggered within the same data collection period and the overall system energy efficiency shows an upward trend, it is determined to be a false prosperity state. The false prosperity state refers to an abnormal operating state in which the overall system energy efficiency continues to rise but the task response delay is simultaneously approaching the service level agreement bottom line, and the energy efficiency index and performance index show a divergence in direction. Understandably, the functional significance of the efficiency and latency monitoring module lies in: by establishing a joint monitoring system for energy efficiency and latency, it enables real-time anomaly identification of system operation status, can capture false prosperity states where energy efficiency indicators and performance indicators diverge, can detect performance degradation risks that may be caused by energy efficiency optimization strategies in advance, and solves the problem that traditional steady-state assessment cannot warn of dynamic anomalies. Optimize the safety definition module: Based on the false prosperity state of efficiency, implement a dynamic load intensity increase strategy, track the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits, and define the effective scope of energy efficiency optimization of energy efficiency scheduling strategy without triggering performance degradation. In the optimized safety definition module, based on the false prosperity state of efficiency, the dynamic load intensity increase strategy is implemented to track the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits as follows: The total system load intensity at the time of triggering the false prosperity state is used as the starting value. The unit of load intensity is the number of concurrent tasks or the percentage of resource utilization. The load intensity is set to increase by 5% of the starting value in each step. For example, if the starting load intensity is 100 concurrent tasks, then the first step increases to 105 concurrent tasks, and the second step increases to 110 concurrent tasks. Each ramp-up includes the following operations: increasing the load intensity by one step; maintaining the load intensity for 60 seconds to allow the system to stabilize; and collecting system performance and energy efficiency metrics during the last 10 seconds of stable operation. The performance metrics are the average of the task response latency of all nodes, and the energy efficiency metrics are the average of the real-time energy efficiency ratio of all nodes. Starting from the initial load intensity, the load intensity is increased round by round. After each round, the average task response latency of the current round is compared with the service level agreement (SLA) minimum value. If the average response latency of the current round is less than the minimum value, the next round of increasing load intensity continues. If the average response latency of the current round reaches or exceeds the minimum value for the first time, the load intensity corresponding to the previous round is recorded as the critical point of system performance collapse. For example, if the response latency exceeds the minimum value for the first time when the load intensity of the 5th round is 120 concurrent tasks, then the load intensity of the 4th round with 115 concurrent tasks is the critical point of performance collapse. During the progressive scaling process, after each round, the system's average energy efficiency ratio (EER) for the current round is recorded and compared with that of the previous round. If the EER of the current round is higher than that of the previous round, the next round of scaling continues. If the EER of the current round first falls below that of the previous round, the load intensity corresponding to the previous round is recorded as the inflection point of diminishing energy efficiency gains. For example, if the EER of the third round is 2.5 and the EER of the fourth round is 2.4, marking the first decrease, then the load intensity of the third round is the inflection point of diminishing energy efficiency gains. An example is shown in Table 4 below. Table 4: Dynamic Increase in Load Intensity; ; In the optimization safety delineation module, the process of defining the effective scope of energy efficiency scheduling strategies without triggering performance degradation is as follows: The smaller value of the load intensity corresponding to the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits is taken as the upper limit of the effective scope of energy efficiency optimization. The initial load intensity at which the false prosperity state of efficiency is triggered is taken as the lower limit of the effective scope, which is the starting load intensity of the dynamic climbing process; thus, the effective scope of energy efficiency optimization is obtained. It should be noted that the effective scope of energy efficiency optimization is a load intensity range. Any load intensity within the load intensity range must meet the following two conditions: the average task response latency does not exceed the service level agreement minimum, and the energy efficiency ratio has not yet started to decline. Executing the energy efficiency scheduling strategy within the load intensity range will not trigger performance degradation. Understandably, the significance of optimizing the safety definition module lies in: by gradually approaching the load intensity through experimental methods, tracking the critical point of performance collapse and the inflection point of diminishing energy efficiency benefits respectively, defining the safe and effective load range of energy efficiency scheduling strategy without triggering performance degradation, providing a clear safety boundary reference for energy efficiency scheduling, and solving the problem of lack of quantitative basis for scheduling decisions; The optimization strategy correction module identifies the failure tendency of existing scheduling strategies that deviate from the effective scope of energy efficiency optimization, and classifies and optimizes existing scheduling strategies: for strategies with delayed response, the multi-dimensional monitoring system is adjusted to detect the critical risk of approaching performance collapse earlier; for strategies with energy efficiency overdraft, the dynamic load game field is adjusted so that scheduling decisions are proactively withdrawn before the inflection point of diminishing energy efficiency benefits is crossed. In the optimization strategy correction module, the process of identifying the failure tendency of existing scheduling strategies that deviate from the effective scope of energy efficiency optimization and classifying existing scheduling strategies is as follows: Obtain the current system load intensity value and compare it with the effective scope of energy efficiency optimization: If the actual load intensity of the system at the current moment is less than the lower limit of the effective range of energy efficiency optimization, it is judged as a deviation tendency of not fully utilizing the safety range; If the actual load intensity of the system at the current moment is greater than the upper limit of the effective scope of energy efficiency optimization, it is determined that the deviation tendency has exceeded the safety boundary. For deviations that do not fully utilize the safety zone, the current scheduling strategy is classified as a response lag strategy. The strategy is characterized by the system prematurely implementing conservative rollback operations before approaching the performance collapse threshold, resulting in the load intensity remaining below the lower bound of the safety zone for a long time, and failing to fully utilize the energy efficiency optimization space within the safety zone. For deviations that have crossed the safety boundary, the current scheduling strategy is classified as an energy efficiency overdraft strategy. The strategy is characterized by the following: after the system has passed the inflection point of diminishing energy efficiency benefits, it continues to concentrate the load on high energy efficiency nodes, causing the load intensity to exceed the upper limit of the safety domain, forcing the system to cross the performance collapse critical point. In the strategy correction module, for strategies with delayed responses, the process of adjusting the multi-dimensional monitoring system to detect the critical risk of performance collapse earlier is as follows: Extract the average task response latency corresponding to multiple alarm triggers from the historical records of the performance latency monitoring module, and take the average of the last five alarm triggers as the current alarm trigger value; For each alarm triggered, trace back from the alarm time to find the first time when the average response latency exceeds the service level agreement threshold, extract the corresponding average response latency as the actual threshold, and take the average of the five actual threshold values ​​triggered by the alarm as the final actual threshold. Calculate the gap between the current alarm trigger value and the actual critical value. If the gap exceeds 10% of the Service Level Agreement (SLA) baseline value, it indicates that the alarm was triggered too early and the scheduling system implemented a premature rollback. In this case, adjust the alarm trigger value (the percentage of the SLA baseline value) of the monitoring system downward from the original setting: adjust it gradually by decreasing it by 5% each time until the gap between the alarm trigger value and the actual critical value is reduced to within 5% of the baseline value; so that it can detect the critical risk of approaching performance collapse earlier; as shown in Table 5 below. Table 5: Comparison of the monitoring system before and after adjustment; ; In the optimization strategy correction module, for energy efficiency overdraft strategies, the dynamic load game field is adjusted so that the scheduling decision actively retracts before crossing the inflection point of diminishing energy efficiency benefits. The process is as follows: Extract the load intensity corresponding to the point when the system load intensity begins to decrease in the current actual operation from the operation record of the dynamic load game field; read the load intensity corresponding to the recorded inflection point of diminishing energy efficiency benefits from the tracking results of the optimized safety demarcation module. Extract the load value corresponding to the point when the system load intensity begins to decrease in the current actual operation from the operation record of the dynamic load game field, and record it as the current pullback point; Read the load value corresponding to the recorded inflection point of diminishing energy efficiency benefits from the tracking results of the optimized safety definition module; If the load value corresponding to the current pullback point is greater than the load value corresponding to the inflection point of diminishing energy efficiency benefits, it means that the system has already passed the inflection point of diminishing energy efficiency benefits when the actual pullback occurs. That is, the main body only started to reduce the amount of resources requested after passing the inflection point. Each time, the energy consumption cost coefficient in the dynamic load game field is increased by 10% from the current value. The game field is then rerun and the new pullback point is observed. If the load value corresponding to the pullback point is still greater than the load value corresponding to the inflection point in the new round of operation, the energy consumption cost coefficient is increased until the load value corresponding to the pullback point is equal to or less than the load value corresponding to the inflection point. This encourages the entity to actively reduce resource requests before the inflection point in subsequent operations. For example, as shown in Table 6 below. Table 6: Comparison of Dynamic Load Game Field Before and After Adjustment; ; Understandably, the significance of the optimization strategy correction module lies in: by identifying the failure tendency of scheduling strategies that deviate from the safe range, classifying the strategies into response lag type and energy efficiency overdraft type, and making parameter corrections for the monitoring system and the game field respectively, it realizes closed-loop governance from anomaly identification to strategy optimization, enabling the scheduling system to have adaptive correction capabilities and continuously maintain a dynamic balance between energy efficiency optimization and performance assurance.

[0021] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A marginal energy efficiency scheduling system for joint game theory among computing servers, characterized in that: Includes the following modules: Load game construction module: By tracking the historical task load sequence of the computing power cluster, extracting load characteristics and identifying the resource demand distribution and conflict relationship of multiple subjects, a dynamic load game field containing load characteristics and subject competition relationship is constructed. Performance and latency monitoring module: Under the operating conditions of dynamic load game field, collect the real-time energy efficiency ratio and task response latency of each node, and establish a multi-dimensional monitoring system for energy efficiency and latency: when the task response latency approaches the service level agreement bottom line while the overall system energy efficiency still shows an upward trend, it is judged as a false prosperity state of performance. Optimize the safety definition module: Based on the false prosperity state of efficiency, implement a dynamic load intensity increase strategy, track the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits, and define the effective scope of energy efficiency optimization of energy efficiency scheduling strategy without triggering performance degradation. The optimization strategy correction module identifies the failure tendency of existing scheduling strategies that deviate from the effective scope of energy efficiency optimization, and classifies and optimizes existing scheduling strategies: for strategies with delayed response, the multi-dimensional monitoring system is adjusted to detect the critical risk of approaching performance collapse earlier; for strategies with energy efficiency overdraft, the dynamic load game field is adjusted so that scheduling decisions are proactively withdrawn before the inflection point of diminishing energy efficiency benefits is crossed.

2. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method of extracting load characteristics and identifying the distribution and conflict relationships of resource demands among multiple stakeholders by tracking the historical task load sequences of computing power clusters is as follows: Export historical task execution records from the scheduling system logs, retaining records of tasks that were completed normally; Calculate task waiting time, runtime, and resource request volume; group by tenant or queue identifier to count the total number of tasks, average and maximum resource request volume, and task arrival time sequence for each subject; divide into hourly segments to count the total resource request volume for each subject; if the total resource request volume exceeds the total capacity of the same type of resources in the cluster, it is determined that there is a competition conflict and the conflicting subject pair and conflict time period are recorded.

3. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method for constructing a dynamic load game field is as follows: Each tenant or queue is defined as a participating entity, which has independent decision-making power for resource requests. The system status is described by the current availability of three types of resources: CPU, memory, and GPU. The current availability is equal to the total resource capacity minus the occupied amount. Each entity submits a resource request based on the current system status in each scheduling time slot. The scheduler aggregates the requests and allocates resources according to the requests. After allocation, the system status is updated. The system status is restored after the task is completed and resources are released.

4. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 3, characterized in that: In a dynamic load game, the utility function of each subject consists of a benefit term and a cost term. The benefit term is calculated by multiplying the square of the ratio of the actual allocated resources to the requested resources by a benchmark value. The larger the allocation gap, the faster the loss of benefits. The cost item includes the sum of completion time cost and energy consumption cost. The completion time cost equals the task runtime multiplied by the time cost coefficient, and the energy consumption cost equals the total energy consumption of the cluster multiplied by the energy consumption cost coefficient and is shared equally by all entities. The utility value equals the benefit item minus the completion time cost and then minus the energy consumption cost.

5. A marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 4, characterized in that: Setting multi-party competition penalty rules in dynamic load game: Using scheduling time slots as the unit window, if multiple entities with conflicting relationships apply for resources in any type of resource and their application amounts all exceed the set percentage of their respective baseline values, a competition penalty will be triggered. After triggering, the energy consumption cost of each triggering entity is increased by an additional penalty value based on the current value. The additional penalty value is equal to the current energy consumption cost multiplied by the penalty coefficient. Each triggering entity calculates and adds the penalty value to its own energy consumption cost independently.

6. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method for collecting real-time energy efficiency ratio and task response latency of each node is as follows: Using each computing server as the data collection object, the instantaneous power consumption of the node is read through the hardware management interface, and the computing power output value of the node is read through the operating system interface. The real-time energy efficiency ratio is equal to the computing power output value divided by the power consumption value. The submission time of all running tasks on the node is obtained from the task queue of the scheduling system, and the response latency is calculated. The average response latency of all running tasks on the node is taken as the representative value of the node's current collection period.

7. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method for establishing a multi-dimensional monitoring system is as follows: The monitoring system includes energy efficiency and latency dimensions. For each node, two time series, energy efficiency and latency, are maintained and updated on a rolling basis. The energy efficiency ratio and response latency values ​​obtained by the same node in the same acquisition cycle are marked as paired data points under the same timestamp. The pairing relationship ensures that changes in energy efficiency and latency can be correlated and compared on the same time base in subsequent analysis.

8. The marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method for determining a false prosperity in efficiency is as follows: After each data collection cycle is completed, the average response delay of all nodes is calculated. If the average response delay reaches the set percentage of the service level agreement's minimum value, the approximation condition is triggered. At the same time, the average system energy efficiency of the current and the previous five data collection cycles is taken to determine the trend. If the slope of the fitted line is greater than zero, the energy efficiency is determined to be on an upward trend. When the approximation condition and the upward trend of energy efficiency are met simultaneously in the same data collection cycle, it is determined to be a false prosperity state of efficiency.

9. A marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The method for tracking the critical point of system performance collapse and the inflection point of diminishing energy efficiency benefits is as follows: Starting with the load intensity at which a false boom state is triggered, the load intensity is increased in increments of a percentage from the initial value. After each round of stable operation, the average response latency and average energy efficiency ratio are collected. The load intensity of the round before the average response latency first reaches or exceeds the service level agreement threshold is recorded as the performance collapse critical point. The point at which the average energy efficiency ratio first falls below the load intensity of the previous round is recorded as the inflection point of decreasing energy efficiency benefits; the smaller of the two corresponding load intensities is taken as the upper limit of the effective scope, and the initial load intensity is taken as the lower limit of the effective scope.

10. A marginal energy efficiency scheduling system for joint game theory of computing servers according to claim 1, characterized in that: The process of dividing and optimizing the existing scheduling strategy is as follows: Compare the current actual load intensity with the effective scope. If the actual load intensity is less than the lower limit of the scope, it is determined to be a response lag strategy; if it is greater than the upper limit of the scope, it is determined to be an energy efficiency overdraft strategy. For response lag strategies, lower the monitoring and alarm threshold so that the alarm trigger value is close to the actual critical value. For energy efficiency overdraft strategies, increase the energy consumption cost coefficient in the dynamic load game field until the load intensity corresponding to the system pullback point is equal to or less than the load intensity corresponding to the energy efficiency reduction inflection point.

Images