A load prediction-based energy-saving management and control method and system for a computing power cluster

Abstract

This invention discloses a method and system for energy-saving management and control of computing power clusters based on load prediction, belonging to the field of energy-saving management and control technology. It collects load utilization data, node energy density data, and task latency adaptation data; constructs a data sampling time period and a set of computing power cluster nodes; constructs a basic state matrix of all computing power cluster nodes at the data sampling time point; calculates the comprehensive load state quantity of the computing power cluster nodes; calculates the maximum adjustable power and the adjustable range of task latency; calculates the comprehensive energy-saving potential value; based on the comprehensive energy-saving potential value and the comprehensive load state quantity, calculates the node load-potential matching degree of the computing power cluster nodes; calculates the node load migration amount of the computing power cluster nodes at the data sampling time point; analyzes and performs energy-saving management and control, ensuring that the energy-saving process takes into account performance constraints, scheduling rationality, and system stability, thereby improving the overall energy efficiency and operational reliability of large-scale computing power clusters.

Description

Technical Field

[0001] This invention relates to the field of energy-saving management and control technology, specifically to a method and system for energy-saving management and control of computing power clusters based on load prediction. Background Technology

[0002] With the rapid development of cloud computing, edge computing, and artificial intelligence inference services, computing infrastructure is gradually evolving from single-server architecture to large-scale computing clusters. Data centers are characterized by high-density deployment, heterogeneous node coexistence, and significant dynamic fluctuations in business load. Against this backdrop, research on optimizing computing resource utilization efficiency and energy efficiency ratio has deepened. Related technologies have evolved from early capacity configuration methods based on static resource planning to dynamic resource scheduling based on virtualization and containerization, and further to energy efficiency-aware scheduling mechanisms combined with power consumption monitoring. Existing technologies typically collect single-dimensional operating parameters such as CPU utilization, memory usage, or node power values ​​to migrate tasks or control the start / stop of the entire machine, aiming to reduce energy consumption losses caused by idle or inefficient operation. Some solutions further introduce load prediction models to estimate the volume of business requests in the future, thereby enabling advance resource scaling. These technologies improve resource waste to some extent, but overall they still focus on a single dimension from either a "load perspective" or a "power consumption perspective," lacking a multi-source collaborative characterization of the computing node's operating status and failing to accurately reflect the true energy-saving adjustment space of nodes under performance constraints.

[0003] However, existing energy efficiency optimization technologies for computing clusters still have several key shortcomings: First, most solutions separate load utilization and energy consumption indicators, failing to establish a structural correlation model between the two. This makes it impossible for scheduling decisions to distinguish the essential differences between "high-load, high-energy-efficiency nodes" and "high-load, low-energy-efficiency nodes," easily leading to insignificant energy-saving effects or even performance degradation. Second, existing load predictions mostly focus on the volume of business requests or the trend of CPU curve changes, ignoring the load coupling relationship between nodes and the characteristics of historical load structure, making it difficult to support coordinated control at the cluster scale. Third, existing energy-saving control methods mainly rely on node frequency reduction, hibernation, or simple task migration, without quantifying the adjustable range of nodes under the condition of meeting task latency constraints. The lack of a unified quantitative basis for "performance-energy consumption-load" leads to a biased, experience-based control process. These problems directly limit the refined energy-saving capabilities of computing clusters in complex business scenarios and make it difficult to form a stable and predictable closed-loop scheduling mechanism for energy-saving control. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for energy-saving management and control of computing power clusters based on load prediction, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A load prediction-based energy-saving management method for computing clusters includes the following steps: Step S1: Construct a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; construct a data sampling time period and a set of computing cluster nodes; Step S2: Construct the operating feature vector of the computing cluster nodes at the data sampling time point, and construct the basic state matrix of all computing cluster nodes at the data sampling time point; Step S3: Based on the basic state matrix, calculate the comprehensive load state quantity of the computing cluster nodes at the data sampling time point; calculate the maximum adjustable power and adjustable task latency range of the computing cluster nodes at the data sampling time point; calculate the comprehensive energy-saving potential value of the computing cluster nodes at the data sampling time point; Step S4: Based on the comprehensive energy-saving potential value and the comprehensive load state quantity, calculate the node load-potential matching degree of the computing cluster nodes at the data sampling time point; calculate the node load migration amount of the computing cluster nodes at the data sampling time point, and analyze and perform energy-saving management.

[0006] As a preferred embodiment of the energy-saving management method for computing power clusters based on load prediction described in this invention, a load perception and acquisition module is constructed. The load perception and acquisition module is used to collect the core operating data of the computing power cluster. The load perception and acquisition module includes a load monitoring sensor, an energy consumption density acquisition unit, and a latency adaptation detection unit. The core operating data includes load utilization data, node energy consumption density data, and task latency adaptation data. The load monitoring sensor is deployed in the core layer of the computing power cluster to collect load utilization data of the computing power cluster; the energy consumption density acquisition unit is deployed in the power supply link of the computing power cluster to obtain node energy consumption density data by monitoring the energy consumption per unit of computing power; the latency adaptation detection unit records the time difference between task transmission and response by sending standard test data packets and calculates task latency adaptation data.

[0007] As a preferred embodiment of the energy-saving management method for computing clusters based on load prediction described in this invention, a data sampling time period is constructed, denoted as... ,in, Let A represent the a-th data sampling time point, and A represent the total number of data sampling time points; the data sampling time points are respectively... The load utilization data, node energy density data, and task latency adaptation data collected below are denoted as follows: and ; Construct a set of computing power cluster nodes, denoted as . ,in, This represents the b-th computing cluster node, where B represents the total number of computing cluster nodes; the data sampling time point The computing power cluster nodes collected below The load utilization data, node energy density data, and task latency adaptation data are normalized and denoted as follows: and .

[0008] As a preferred embodiment of the energy-saving management method for computing clusters based on load prediction described in this invention, a data sampling time point is constructed. Lower computing power cluster nodes The running feature vector, denoted as ; Based on data sampling time points Lower computing power cluster nodes Operating feature vector Obtain data sampling time points The runtime feature vectors of all computing cluster nodes are obtained, and data sampling time points are constructed. The basic state matrix of all computing cluster nodes is denoted as follows: .

[0009] As a preferred embodiment of the energy-saving management method for computing clusters based on load prediction described in this invention, based on data sampling time points... The basic state matrix of all computing power cluster nodes Calculate the computing cluster nodes at all data sampling time points. The average load utilization data is calculated using the following formula: ,in, This represents the computing cluster nodes at all data sampling time points. The average load utilization data; In this invention, load prediction is manifested as a trend-based assessment and forward-looking regulation based on historical operating data. Specifically, it involves continuously collecting and analyzing historical load utilization data of each computing cluster node to calculate its average over a period of time. This mean not only reflects the node's past load level, but more importantly, it serves as a benchmark or steady-state trend characterization for the node's expected load in the next stage, providing a quantitative basis for predicting its short-term load status. This invention utilizes this predictive benchmark ( By combining data with multi-dimensional dynamic indicators such as energy density, task latency, and load correlation between nodes, a comprehensive load state quantity is constructed that reflects the expected load pressure and operational status of nodes. Based on this, by evaluating the energy-saving potential of nodes and calculating the load-potential matching degree, proactive energy-saving management and dynamic load migration based on predictive load state are ultimately achieved. This method overcomes the lag of relying solely on instantaneous load for decision-making, enabling the system to intelligently intervene before or at the initial stage of load fluctuations, thereby achieving a better energy efficiency balance and cluster stability.

[0010] Calculate data sampling time points The load correlation coefficient between nodes in the computing power cluster is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes With computing power cluster nodes The load correlation coefficient between them Describing covariance, Indicates standard deviation, Indicates the data sampling time point The computing power cluster nodes collected below The load utilization data, and k≠b; Calculate data sampling time points Lower computing power cluster nodes The comprehensive load state quantity is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The overall load state quantity, This represents the preset load impact factor. This represents the preset load correlation coefficient influence factor. This represents the preset node energy consumption density influence factor.

[0011] As a preferred embodiment of the energy-saving management method for computing clusters based on load prediction described in this invention, the data sampling time point is calculated. Lower computing power cluster nodes The maximum adjustable power is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes Maximum adjustable power; Calculate data sampling time points Lower computing power cluster nodes The adjustable range of task delay is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes The adjustable range of task delay. This indicates the maximum allowed task latency threshold for the preset node; Based on data sampling time points Lower computing power cluster nodes Maximum adjustable power and data sampling time points Lower computing power cluster nodes Adjustable task delay range Calculate the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value, This represents the preset maximum adjustable power influence factor. This indicates the preset adjustable range of task delay influence factor.

[0012] As a preferred embodiment of the energy-saving management method for computing clusters based on load prediction described in this invention, based on data sampling time points... Lower computing power cluster nodes Comprehensive energy-saving potential value With data sampling time point Lower computing power cluster nodes Comprehensive load state quantity Calculate the data sampling time point Lower computing power cluster nodes The node load-potential matching degree is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes Node load-potential matching degree; Calculate data sampling time points Lower computing power cluster nodes The node load migration amount is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The amount of node load migration, Indicates the data sampling time point Lower computing power cluster nodes The node load-potential matching degree, and k≠b; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes Can receive load; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes The load needs to be moved out; The system can acquire the node load migration amount of each computing cluster node at the current data sampling time point in real time and perform intelligent energy-saving management.

[0013] A computing power cluster energy-saving management and control system based on load prediction. The system includes: a data acquisition and set construction module, a vector construction and matrix construction module, a state variable and potential value calculation module, and a matching degree calculation and control module. The data acquisition and aggregation construction module includes: constructing a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; and constructing a data sampling time period and a computing cluster node set. The vector construction and matrix construction module: constructs the running feature vector of the computing power cluster node at the data sampling time point, and constructs the basic state matrix of all computing power cluster nodes at the data sampling time point; The state quantity and potential value calculation module calculates the comprehensive load state quantity of the computing power cluster node at the data sampling time point based on the basic state matrix; calculates the maximum adjustable power and adjustable task delay range of the computing power cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing power cluster node at the data sampling time point. The matching degree calculation and control module calculates the node load-potential matching degree of the computing cluster nodes at the data sampling time point based on the comprehensive energy-saving potential value and the comprehensive load status value; it also calculates the node load migration amount of the computing cluster nodes at the data sampling time point, and analyzes and performs energy-saving control.

[0014] Furthermore, the state quantity and potential value calculation module includes a state quantity calculation unit and a potential value calculation unit; The state quantity calculation unit: based on the basic state matrix of all computing power cluster nodes at the data sampling time point, calculates the average load utilization data of the computing power cluster nodes at all data sampling time points; calculates the load correlation coefficient between two adjacent computing power cluster nodes at the data sampling time point; and calculates the comprehensive load state quantity of the computing power cluster nodes at the data sampling time point. The potential value calculation unit: calculates the maximum adjustable power of the computing cluster node at the data sampling time point; calculates the adjustable range of the task delay of the computing cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing cluster node at the data sampling time point based on the maximum adjustable power and the adjustable range of the task delay of the computing cluster node at the data sampling time point.

[0015] Furthermore, the matching degree calculation and control module includes a matching degree calculation and control unit; The matching degree calculation and control unit calculates the node load-potential matching degree of the computing cluster nodes at the data sampling time point based on the comprehensive energy-saving potential value and comprehensive load status of the computing cluster nodes at the data sampling time point. It also calculates the node load migration amount of the computing cluster nodes at the data sampling time point. If the node load migration amount of the computing cluster nodes at the data sampling time point is greater than zero, it is determined that the computing cluster nodes at the data sampling time point can accept the load; if the node load migration amount of the computing cluster nodes at the data sampling time point is less than zero, it is determined that the computing cluster nodes at the data sampling time point need to migrate the load. It also obtains the node load migration amount of each computing cluster node at the current data sampling time point in real time and performs intelligent energy-saving control.

[0016] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention provides a method and system for energy-saving management of computing clusters based on load prediction. By constructing a load perception and acquisition module, it synchronously acquires load utilization, node energy consumption density, and task latency adaptation data, and establishes a unified data sampling time system and node identification system. This achieves homogeneous perception modeling of the computing cluster in three dimensions: computational pressure, energy consumption characteristics, and performance tolerance boundaries. From the source, this enables energy-saving control to have business constraint recognition capabilities, avoiding performance degradation caused by adjusting solely based on load changes. Furthermore, by mapping multi-source operational data to node operational feature vectors and constructing a basic state matrix, it realizes the transformation from discrete monitoring data to a structured cluster state space, making node states computable and comparable, providing a unified mathematical carrier for subsequent pattern modeling. Further, on the one hand, it utilizes historical negative data... By constructing a comprehensive load state variable based on the load average and the load correlation between nodes, a regular characterization of the node load structure and energy consumption characteristics is achieved, elevating load prediction from instantaneous monitoring to structural modeling. On the other hand, a comprehensive energy-saving potential value is calculated by combining the maximum adjustable power and the adjustable range of task latency, quantifying the energy-saving space of nodes in the current scenario by combining power regulation margin with the service's sacrificial performance boundary. This allows energy-saving decisions to be based on adjustable capacity awareness rather than static power consumption judgment. Furthermore, a node load-potential matching degree is constructed using the comprehensive load state variable and the comprehensive energy-saving potential value, achieving a unified characterization of the balance between load pressure and energy-saving space. The node load migration amount is further calculated based on the difference between the matching degree and the global level, structurally reconstructing the cluster load distribution. This reduces the load of high-energy-consuming nodes and increases the utilization rate of low-load nodes, thereby optimizing the energy consumption distribution without weakening computing power. Through the above technical path, this invention forms an energy-saving management mechanism driven by load prediction and using energy efficiency structure rearrangement as a means, ensuring that the energy-saving process takes into account performance constraints, scheduling rationality, and system stability, improving the overall energy efficiency and operational reliability of large-scale computing clusters. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a schematic diagram illustrating the steps of a computing cluster energy-saving management method based on load prediction according to the present invention. Figure 2 This is a schematic diagram of the structure of a computing power cluster energy-saving management and control system based on load prediction according to the present invention. Detailed Implementation

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

[0020] Please see Figure 1 In this first embodiment, a method for energy-saving management and control of computing clusters based on load prediction is provided. The method includes the following steps: Step S1: Construct a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; construct the data sampling time period and the set of computing cluster nodes.

[0021] Specifically, a load-aware acquisition module is constructed. The load-aware acquisition module is used to collect the core operating data of the computing power cluster. The load-aware acquisition module includes a load monitoring sensor, an energy consumption density acquisition unit, and a latency adaptation detection unit. The core operating data includes load utilization data, node energy consumption density data, and task latency adaptation data. The load monitoring sensor is deployed in the core layer of the computing power cluster to collect load utilization data of the computing power cluster; the energy consumption density acquisition unit is deployed in the power supply link of the computing power cluster to obtain node energy consumption density data by monitoring the energy consumption per unit of computing power; the latency adaptation detection unit records the time difference between task transmission and response by sending standard test data packets and calculates task latency adaptation data.

[0022] Furthermore, the data sampling time period is constructed, denoted as . ,in, Let A represent the a-th data sampling time point, and A represent the total number of data sampling time points; the data sampling time points are respectively... The load utilization data, node energy density data, and task latency adaptation data collected below are denoted as follows: and ; Construct a set of computing power cluster nodes, denoted as . ,in, This represents the b-th computing cluster node, where B represents the total number of computing cluster nodes; the data sampling time point The computing power cluster nodes collected below The load utilization data, node energy density data, and task latency adaptation data are normalized and denoted as follows: and .

[0023] In this invention, by constructing a load perception acquisition module, load utilization, node energy consumption density and task latency adaptation data are collected synchronously, and a unified data sampling time period and node set identification system are established, thus realizing a multi-dimensional and homogeneous perception modeling foundation for the computing power cluster's operating status.

[0024] Load monitoring sensors acquire computing power consumption intensity from the computing core layer, reflecting the "computing power pressure status"; energy consumption density acquisition units establish a "unit computing power energy consumption model" from the power supply link side, enabling a direct correspondence between energy consumption and computing behavior; delay adaptation detection units establish "performance constraint boundaries" from the service response layer.

[0025] This step enables a unified three-dimensional characterization of load, energy consumption, and performance tolerance, allowing subsequent energy-saving control to no longer rely solely on load data but to possess the ability to recognize energy efficiency constraints. This avoids the problem of "reducing load to save energy but degrading business performance" in traditional methods from the source, and improves the controllability and engineering applicability of energy-saving decisions.

[0026] Step S2: Construct the running feature vector of the computing cluster nodes at the data sampling time point, and construct the basic state matrix of all computing cluster nodes at the data sampling time point.

[0027] Specifically, constructing data sampling time points Lower computing power cluster nodes The running feature vector, denoted as ; Based on data sampling time points Lower computing power cluster nodes Operating feature vector Obtain data sampling time points The runtime feature vectors of all computing cluster nodes are obtained, and data sampling time points are constructed. The basic state matrix of all computing cluster nodes is denoted as follows: .

[0028] In this invention, by constructing node running feature vectors at time points and forming a basic state matrix, the transformation from "discrete sensing data" to "structured cluster state space" is realized.

[0029] This step transforms data from different physical sources into feature vectors of a unified dimension; it explicitly expresses the two-dimensional coupling relationship between nodes and time through a matrix structure. This structure makes the cluster's operating state computable and comparable, providing a mathematical foundation for subsequent load correlation modeling and energy-saving potential estimation, transforming energy-saving decisions from "empirical rules" to "state-space operations," significantly enhancing the algorithm's stability and scalability.

[0030] Step S3: Based on the basic state matrix, calculate the comprehensive load state of the computing cluster nodes at the data sampling time point; calculate the maximum adjustable power and adjustable task delay range of the computing cluster nodes at the data sampling time point; calculate the comprehensive energy-saving potential value of the computing cluster nodes at the data sampling time point.

[0031] Specifically, based on the data sampling time point The basic state matrix of all computing power cluster nodes Calculate the computing cluster nodes at all data sampling time points. The average load utilization data is calculated using the following formula: ,in, This represents the computing cluster nodes at all data sampling time points. The average load utilization data; Calculate data sampling time points The load correlation coefficient between nodes in the computing power cluster is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes With computing power cluster nodes The load correlation coefficient between them Describing covariance, Indicates standard deviation, Indicates the data sampling time point The computing power cluster nodes collected below The load utilization data, and k≠b; It should be noted that covariance reflects the performance of the computing cluster nodes. With computing power cluster nodes The load change trend is consistent (positive covariance indicates the same trend, negative covariance indicates the opposite trend); the standard deviation reflects the fluctuation range of the load of a single node; existing technologies ignore the load correlation between nodes, resulting in isolated scheduling decisions (such as only migrating the load of a single node, causing overload of associated nodes), this coefficient can identify node groups with "synchronous load fluctuations"; it provides a "cluster perspective" for comprehensive load state variables: avoiding focusing only on the load of a single node while ignoring its collaborative relationship with other nodes.

[0032] like =0.9: Computing power cluster node With computing power cluster nodes The load is highly correlated (e.g., all are undertaking different sub-tasks of the same business), so migrate computing cluster nodes. When the load is high, avoid migrating to the computing cluster node. Otherwise, it will cause the computing power cluster nodes to... Overload; like =−0.7: Computing power cluster node With computing power cluster nodes The load is negatively correlated (such as the computing power cluster nodes). Peak computing power cluster nodes During off-peak hours, computing cluster nodes can be prioritized. The load is migrated to the computing cluster nodes. This enables cluster load balancing.

[0033] Calculate data sampling time points Lower computing power cluster nodes The comprehensive load state quantity is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The overall load state quantity, This represents the preset load impact factor. This represents the preset load correlation coefficient influence factor. This represents the preset node energy consumption density influence factor.

[0034] For "latency-sensitive businesses (such as AI inference)," it can improve (Avoid conflicts between related nodes) and (Prioritize low-energy nodes); for "throughput-priority services (such as data backup)," efficiency can be improved. (Focusing on load balancing), among which, , and All presets are based on historical data and expert experience; It should be noted that, based on the concept of multi-dimensional collaborative modeling, three types of key information are integrated: the node's own long-term load level ( ), the load correlation between the node and other nodes in the cluster ( That is, the average correlation coefficient), and the node's own energy consumption characteristics ( The higher the energy density, the more "power-consuming" the node becomes.

[0035] Breaking through the limitations of existing technologies that rely on "single-dimensional load assessment": Current technologies only use CPU utilization to determine load, while this formula achieves a three-dimensional load status characterization encompassing "load level + cluster connectivity + energy consumption characteristics." It provides a "basic state benchmark" for subsequent energy-saving potential calculations: the higher the overall load status, the greater the current load pressure on the node, the tighter the connectivity, and the higher the energy consumption, resulting in lower energy-saving potential; conversely, the lower the overall load status, the greater the energy-saving potential.

[0036] Assumption: =0.4 (load weight) =0.3 (association weight) =0.3 (energy consumption weight); computing power cluster nodes of =0.8 (long-term high load) =0.7 (highly correlated with other nodes) =0.9 (high energy density); Calculated This indicates that the node is in a state of "high load, high correlation, and high energy consumption" with very low energy-saving potential. The load should not be forcibly migrated (otherwise it will affect business continuity).

[0037] In this invention, a comprehensive load state quantity is constructed by combining historical load averages with inter-node load correlation coefficients, achieving structural predictive modeling of node operating load rather than instantaneous monitoring. The load average characterizes the long-term load level of nodes, reflecting the resource occupancy baseline; the load correlation coefficient reveals the cooperative or competitive relationships between nodes, reflecting scheduling coupling risks; and energy density directly correlates load state with energy characteristics, achieving a fusion of load pattern modeling and energy efficiency correlation modeling. This avoids the problem of traditional scheduling only considering CPU utilization while ignoring node energy consumption differences, thus improving the engineering effectiveness of load prediction.

[0038] Furthermore, calculate the data sampling time points. Lower computing power cluster nodes The maximum adjustable power is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes Maximum adjustable power; Calculate data sampling time points Lower computing power cluster nodes The adjustable range of task delay is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes The adjustable range of task delay. This indicates the maximum allowed task latency threshold for the preset node; Based on data sampling time points Lower computing power cluster nodes Maximum adjustable power and data sampling time points Lower computing power cluster nodes Adjustable task delay range Calculate the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value, This represents the preset maximum adjustable power influence factor. This indicates the preset adjustable range of task delay influence factor.

[0039] For energy-sensitive scenarios (such as edge data centers), improve (Prioritize reducing power); for latency-sensitive scenarios (such as financial transactions), increase power consumption. (Strictly control the increase in delay), among which, and All presets are based on historical data and expert experience; It should be noted that the energy-saving potential of integrated hardware (maximum adjustable power) ) and performance energy-saving potential (adjustable delay range) ), construct a comprehensive energy-saving potential assessment index for nodes, and achieve a balanced quantification of "energy consumption-performance".

[0040] Breaking through the limitations of existing technologies' "single-dimensional energy-saving judgment": Existing technologies only focus on energy consumption or load, while this formula considers both "reducible power" and "transferable latency" to avoid "sacrificing performance for energy saving" or "giving up energy saving for performance"; it provides a "potential ranking basis" for subsequent load migration decisions: the higher the comprehensive energy-saving potential value, the more suitable the node is as an "energy-saving control object" (more energy consumption can be released through adjustment without affecting business).

[0041] For example, business scenarios: latency-sensitive AI inference ( =0.6, =0.4); Node 1: =0.5 (can be reduced by 50W), =0.1 (can increase latency by 10ms) → =0.6×0.5+0.4×0.1=0.34; Node 2: =0.4 (can be reduced by 40W), =0.3 (can increase latency by 30ms) → =0.6×0.4+0.4×0.3=0.36; Node 2 has higher overall energy-saving potential, so energy-saving regulation of Node 2 should be prioritized (although the power reduction is slightly lower, the delay concession space is larger, which is more in line with the balance needs of delay-sensitive services).

[0042] In this invention, a comprehensive energy-saving potential value is constructed by combining the maximum adjustable power and the adjustable range of task delay, thereby achieving a quantitative assessment of the node's "energy-saving space". The maximum adjustable power reflects the node's power regulation margin under the current load; the adjustable delay range reflects the boundary of service performance that can be sacrificed. The resulting energy-saving potential value is not the static energy consumption, but rather an adjustable energy efficiency capacity indicator.

[0043] This indicator links energy-saving decisions with business tolerance, giving energy-saving control performance guarantees and achieving "controllable energy saving," thus solving the problem that energy-saving strategies in existing technologies are not quantifiable or predictable.

[0044] Step S4: Based on the comprehensive energy-saving potential value and the comprehensive load status, calculate the node load-potential matching degree of the computing power cluster nodes at the data sampling time point; calculate the node load migration amount of the computing power cluster nodes at the data sampling time point, analyze and perform energy-saving management.

[0045] Specifically, based on the data sampling time point Lower computing power cluster nodes Comprehensive energy-saving potential value With data sampling time point Lower computing power cluster nodes Comprehensive load state quantity Calculate the data sampling time point Lower computing power cluster nodes The node load-potential matching degree is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes Node load-potential matching degree; Calculate data sampling time points Lower computing power cluster nodes The node load migration amount is calculated using the following formula: ; in, Indicates the data sampling time point Lower computing power cluster nodes The amount of node load migration, Indicates the data sampling time point Lower computing power cluster nodes The node load-potential matching degree, and k≠b; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes Can receive load; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes The load needs to be moved out; The system can acquire the node load migration amount of each computing cluster node at the current data sampling time point in real time and perform intelligent energy-saving management.

[0046] In this invention, the load-potential matching degree of nodes is calculated by combining the comprehensive energy-saving potential value and the comprehensive load state quantity, thereby realizing a mathematical characterization of the balance relationship between "load pressure" and "energy-saving space". A scheduling decision basis index is constructed, which simultaneously considers: load intensity, energy-saving space, and node energy efficiency differences, realizing the transformation from "single-variable scheduling" to "multi-constraint balanced scheduling", avoiding the incorrect downloading of nodes with high load but no energy-saving space, and improving the rationality of scheduling and energy-saving stability.

[0047] By comparing the matching degree with the global average value to calculate the node load migration amount, adaptive reconstruction of the cluster load distribution is achieved. Positive value nodes absorb the load and negative value nodes release the load, which is equivalent to reshaping the energy consumption distribution pattern of the cluster, so that the load of high energy-consuming nodes decreases and the utilization rate of low energy-consuming nodes increases. Instead of directly reducing computing power, energy saving is achieved through structural rearrangement, reducing the risk of high energy-consuming nodes running at full load continuously, improving the overall energy efficiency balance of the cluster, and thus achieving the prediction-driven energy-saving management goal.

[0048] Please see Figure 2 In this second embodiment: a computing power cluster energy-saving management and control system based on load prediction is provided. The system includes: a data acquisition and set construction module, a vector construction and matrix construction module, a state quantity and potential value calculation module, and a matching degree calculation and control module. The data acquisition and aggregation construction module includes: constructing a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; and constructing a data sampling time period and a computing cluster node set. The vector construction and matrix construction module: constructs the running feature vector of the computing power cluster node at the data sampling time point, and constructs the basic state matrix of all computing power cluster nodes at the data sampling time point; The state quantity and potential value calculation module calculates the comprehensive load state quantity of the computing power cluster node at the data sampling time point based on the basic state matrix; calculates the maximum adjustable power and adjustable task delay range of the computing power cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing power cluster node at the data sampling time point. The matching degree calculation and control module calculates the node load-potential matching degree of the computing cluster nodes at the data sampling time point based on the comprehensive energy-saving potential value and the comprehensive load status value; it also calculates the node load migration amount of the computing cluster nodes at the data sampling time point, and analyzes and performs energy-saving control.

[0049] Furthermore, the state quantity and potential value calculation module includes a state quantity calculation unit and a potential value calculation unit; The state quantity calculation unit: based on the basic state matrix of all computing power cluster nodes at the data sampling time point, calculates the average load utilization data of the computing power cluster nodes at all data sampling time points; calculates the load correlation coefficient between two adjacent computing power cluster nodes at the data sampling time point; and calculates the comprehensive load state quantity of the computing power cluster nodes at the data sampling time point. The potential value calculation unit: calculates the maximum adjustable power of the computing cluster node at the data sampling time point; calculates the adjustable range of the task delay of the computing cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing cluster node at the data sampling time point based on the maximum adjustable power and the adjustable range of the task delay of the computing cluster node at the data sampling time point.

[0050] Furthermore, the matching degree calculation and control module includes a matching degree calculation and control unit; The matching degree calculation and control unit calculates the node load-potential matching degree of the computing cluster nodes at the data sampling time point based on the comprehensive energy-saving potential value and comprehensive load status of the computing cluster nodes at the data sampling time point. It also calculates the node load migration amount of the computing cluster nodes at the data sampling time point. If the node load migration amount of the computing cluster nodes at the data sampling time point is greater than zero, it is determined that the computing cluster nodes at the data sampling time point can accept the load; if the node load migration amount of the computing cluster nodes at the data sampling time point is less than zero, it is determined that the computing cluster nodes at the data sampling time point need to migrate the load. It also obtains the node load migration amount of each computing cluster node at the current data sampling time point in real time and performs intelligent energy-saving control.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load prediction-based energy-saving management method for a computing power cluster, characterized in that, The method includes the following steps: Step S1: Construct a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; construct the data sampling time period and the set of computing cluster nodes; Step S2: Construct the running feature vector of the computing power cluster nodes at the data sampling time point, and construct the basic state matrix of all computing power cluster nodes at the data sampling time point; Step S3: Based on the basic state matrix, calculate the comprehensive load state of the computing cluster nodes at the data sampling time point; calculate the maximum adjustable power and adjustable task delay range of the computing cluster nodes at the data sampling time point; calculate the comprehensive energy-saving potential value of the computing cluster nodes at the data sampling time point. Step S4: Based on the comprehensive energy-saving potential value and the comprehensive load status, calculate the node load-potential matching degree of the computing power cluster nodes at the data sampling time point; calculate the node load migration amount of the computing power cluster nodes at the data sampling time point, analyze and perform energy-saving management.

2. The load prediction-based energy-saving management method for a computing cluster according to claim 1, wherein, The specific implementation process of step S1 includes: A load-aware acquisition module is constructed. The load-aware acquisition module is used to collect the core operating data of the computing power cluster. The load-aware acquisition module includes a load monitoring sensor, an energy consumption density acquisition unit, and a latency adaptation detection unit. The core operating data includes load utilization data, node energy consumption density data, and task latency adaptation data. The load monitoring sensor is deployed in the core layer of the computing power cluster to collect load utilization data of the computing power cluster; the energy consumption density acquisition unit is deployed in the power supply link of the computing power cluster to obtain node energy consumption density data by monitoring the energy consumption per unit of computing power; the latency adaptation detection unit records the time difference between task transmission and response by sending standard test data packets and calculates task latency adaptation data.

3. The load prediction-based energy-saving management method for a computing cluster according to claim 2, characterized in that, The specific implementation process of step S1 also includes: The data sampling time period is constructed, denoted as wherein, represents the a-th data sampling time point, and A represents the total number of data sampling time points; the data sampling time points are respectively denoted as and ; Construct a set of computing power cluster nodes, denoted as ,in, This represents the b-th computing cluster node, where B represents the total number of computing cluster nodes; the data sampling time point The computing power cluster nodes collected below The load utilization data, node energy density data, and task latency adaptation data are normalized and denoted as follows: and .

4. The load prediction-based energy-saving management method for a computing cluster according to claim 3, wherein, The specific implementation process of step S2 includes: Constructing data sampling time points Downward force cluster nodes The running feature vector of the first neural network is denoted as ; Based on data sampling time points Downforce cluster nodes Running feature vectors , obtain the running feature vectors of all downforce cluster nodes at the data sampling time points , and construct the basic state matrix of all downforce cluster nodes at the data sampling time points , denoted as .

5. The load prediction-based energy-saving management method for a computing cluster according to claim 4, wherein, The specific implementation process of step S3 includes: Based on data sampling time points The basic state matrix of all computing power cluster nodes The load utilization rate data mean of all computing power cluster nodes at all data sampling time points is calculated, and the calculation formula is: Wherein, The load utilization rate data mean of all computing power cluster nodes at all data sampling time points is represented. Calculate data sampling time points The load correlation coefficient between nodes in the computing power cluster is calculated using the following formula: ； in, Indicates the data sampling time point Lower computing power cluster nodes With computing power cluster nodes The load correlation coefficient between them Describing covariance, Indicates standard deviation, Indicates the data sampling time point The computing power cluster nodes collected below The load utilization data, and k≠b; Calculate data sampling time points Lower computing power cluster nodes The comprehensive load state quantity is calculated using the following formula: ； in, Indicates the data sampling time point Lower computing power cluster nodes The overall load state quantity, This represents the preset load impact factor. This represents the preset load correlation coefficient influence factor. This represents the preset node energy consumption density influence factor.

6. The energy-saving management method for a computing cluster based on load prediction according to claim 5, characterized in that, The specific implementation process of step S3 also includes: Calculate data sampling time points Lower computing power cluster nodes The maximum adjustable power is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes Maximum adjustable power; Calculate data sampling time points Lower computing power cluster nodes The adjustable range of task delay is calculated using the following formula: ,in, Indicates the data sampling time point Lower computing power cluster nodes The adjustable range of task delay. This indicates the maximum allowed task latency threshold for the preset node; Based on data sampling time points Lower computing power cluster nodes Maximum adjustable power and data sampling time points Lower computing power cluster nodes Adjustable task delay range Calculate the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value is calculated using the following formula: ； in, Indicates the data sampling time point Lower computing power cluster nodes The comprehensive energy-saving potential value, This represents the preset maximum adjustable power influence factor. This indicates the preset adjustable range of task delay influence factor.

7. The energy-saving management method for a computing cluster based on load prediction according to claim 6, characterized in that, The specific implementation process of step S4 includes: Based on data sampling time points Lower computing power cluster nodes Comprehensive energy-saving potential value With data sampling time point Lower computing power cluster nodes Comprehensive load state quantity Calculate the data sampling time point Lower computing power cluster nodes The node load-potential matching degree is calculated using the following formula: ； in, Indicates the data sampling time point Lower computing power cluster nodes Node load-potential matching degree; Calculate data sampling time points Lower computing power cluster nodes The node load migration amount is calculated using the following formula: ； in, Indicates the data sampling time point Lower computing power cluster nodes The amount of node load migration, Indicates the data sampling time point Lower computing power cluster nodes The node load-potential matching degree, and k≠b; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes Can receive load; If the data sampling time point Lower computing power cluster nodes Node load migration Then determine the data sampling time point. Lower computing power cluster nodes The load needs to be moved out; The system can acquire the node load migration amount of each computing cluster node at the current data sampling time point in real time and perform intelligent energy-saving management.

8. A load-prediction-based energy-saving management system for computing clusters, executing the load-prediction-based energy-saving management method for computing clusters as described in any one of claims 1-7, characterized in that, The system includes: a data acquisition and set construction module, a vector construction and matrix construction module, a state quantity and potential value calculation module, and a matching degree calculation and control module; The data acquisition and aggregation construction module includes: constructing a load-aware acquisition module to collect load utilization data, node energy density data, and task latency adaptation data; and constructing a data sampling time period and a computing cluster node set. The vector construction and matrix construction module: constructs the running feature vector of the computing power cluster node at the data sampling time point, and constructs the basic state matrix of all computing power cluster nodes at the data sampling time point; The state quantity and potential value calculation module calculates the comprehensive load state quantity of the computing power cluster node at the data sampling time point based on the basic state matrix; calculates the maximum adjustable power and adjustable task delay range of the computing power cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing power cluster node at the data sampling time point. The matching degree calculation and control module calculates the node load-potential matching degree of the computing cluster nodes at the data sampling time point based on the comprehensive energy-saving potential value and the comprehensive load status value; it also calculates the node load migration amount of the computing cluster nodes at the data sampling time point, and analyzes and performs energy-saving control.

9. A computing cluster energy-saving management and control system based on load prediction according to claim 8, characterized in that: The state quantity and potential value calculation module includes a state quantity calculation unit and a potential value calculation unit; The state quantity calculation unit calculates the average load utilization data of all computing cluster nodes at all data sampling time points based on the basic state matrix of all computing cluster nodes at the data sampling time points. Calculate the load correlation coefficient between two adjacent computing cluster nodes at the data sampling time point; Calculate the overall load status of the computing cluster nodes at the data sampling time point; The potential value calculation unit: calculates the maximum adjustable power of the computing cluster node at the data sampling time point; calculates the adjustable range of the task delay of the computing cluster node at the data sampling time point; and calculates the comprehensive energy-saving potential value of the computing cluster node at the data sampling time point based on the maximum adjustable power and the adjustable range of the task delay of the computing cluster node at the data sampling time point.

10. A computing cluster energy-saving management and control system based on load prediction according to claim 9, characterized in that: The matching degree calculation and control module includes a matching degree calculation and control unit; The matching degree calculation and control unit: Based on the comprehensive energy-saving potential value and comprehensive load status of the computing power cluster nodes at the data sampling time point, calculates the node load-potential matching degree of the computing power cluster nodes at the data sampling time point, calculates the node load migration amount of the computing power cluster nodes at the data sampling time point, and if the node load migration amount of the computing power cluster nodes at the data sampling time point is greater than zero, it is determined that the computing power cluster nodes at the data sampling time point can accept the load. If the node load migration amount of the computing power cluster node is less than zero at the data sampling time point, it is determined that the computing power cluster node needs to migrate out of the load at the data sampling time point. The system can acquire the node load migration amount of each computing cluster node at the current data sampling time point in real time and perform intelligent energy-saving management.

Images