An energy efficiency-aware scheduling method and system based on HTCondor and fusing heat dissipation loss

Abstract

This invention relates to an energy efficiency-aware scheduling method and system based on HTCondor with integrated heat dissipation loss, belonging to the field of computer technology. By constructing a multi-dimensional energy efficiency evaluation model for nodes and deeply integrating it into the HTCondor scheduling process, the method prioritizes job assignment to high-energy-efficiency nodes, improves the overall computing power energy efficiency ratio of the computing cluster, and achieves energy saving and consumption reduction. Furthermore, considering the physical characteristic that node fan power consumption increases non-linearly with system load, a heat dissipation loss penalty operator is constructed and introduced, compensating for the shortcomings of traditional scheduling methods that do not consider the impact of heat dissipation load on the total energy efficiency of nodes, making scheduling decisions more consistent with actual physical energy efficiency operation patterns.

Description

Technical Field

[0001] This invention relates to an energy efficiency-aware scheduling method and system based on HTCondor for integrated heat dissipation loss, belonging to the field of computer technology. Background Technology

[0002] Currently, energy conservation and energy efficiency optimization in high-performance computing clusters and data centers have become core development directions in the industry. How to reduce the overall energy consumption of computing clusters and improve the energy efficiency ratio of computing power is a key technical problem that needs to be solved in the field of data center operation and resource scheduling. HTCondor, as a mainstream open-source distributed resource scheduling framework, relies on the ClassAd declarative specification to match resource supply with task requirements, possessing excellent heterogeneous adaptability and fault tolerance, and is therefore widely used in job management and resource scheduling of large-scale computing clusters. However, in complex heterogeneous computing environments, HTCondor's default resource matching logic has significant limitations in environmental awareness; its decision-making process highly depends on the real-time availability of hardware nominal attributes such as the number of CPU cores and physical memory capacity of computing nodes.

[0003] Under this mechanism, the scheduler, by acquiring node resource status, drives job flow to nodes or slots with ample resource reserves, aiming to maximize the nominal space occupancy of physical nodes / slots. However, this space-oriented matching mode is essentially an energy-inefficiency-aware scheduling model, unable to effectively characterize the dynamic energy efficiency differences between heterogeneous nodes executing computing tasks. Furthermore, existing energy-saving scheduling schemes often focus on node utilization or single power consumption indicators, failing to incorporate dynamic heat dissipation losses into the core evaluation dimension. In real-world scenarios, this lack of energy efficiency awareness can cause deviations in the scheduler's RANK matching logic. The scheduling system tends to allocate a large number of jobs to low-energy-efficiency nodes that meet the job requirements but have high heat dissipation costs or high base power consumption, while high-energy-efficiency nodes with higher computing power-to-power ratios are often in an unsaturated state because they are not assigned higher matching weights. This prevents the system from achieving optimal resource scheduling from a global energy consumption perspective, and to some extent exacerbates hardware fatigue wear and the risk of localized heat accumulation on some low-energy-efficiency nodes. Therefore, this invention is proposed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an energy efficiency-aware scheduling method and system based on HTCondor, incorporating heat dissipation loss. It aims to solve the core technical problem in HTCondor-based computing cluster scheduling scenarios where existing conventional configuration modes lack sufficient awareness of node dynamic energy efficiency characteristics, leading to global energy consumption imbalances in the cluster. Existing RANK matching scheduling mechanisms primarily focus on the dynamic redundancy of nominal resources such as CPU cores and memory. This has certain limitations in heterogeneous computing environments. Because it fails to effectively quantify the coupling relationship between real-time computing power, power consumption, and heat dissipation loss of different physical nodes (such as new and old devices, devices of different brands / configurations) during task execution, the scheduling system struggles to identify differences in node dynamic energy efficiency. This invention constructs a multi-dimensional energy efficiency evaluation model and deeply integrates it into the HTCondor scheduling process, enabling priority assignment of jobs to high-energy-efficiency nodes, thereby improving the overall computing power-energy efficiency ratio of the computing cluster.

[0005] This invention simultaneously addresses the following secondary technical issues, further optimizing scheduling adaptability and cluster operational stability: 1. Nonlinear characterization of heat dissipation loss of heterogeneous nodes: In view of the physical characteristic that the power consumption of node fans increases nonlinearly with the increase of system load, a heat dissipation loss penalty operator is constructed and introduced to make up for the shortcomings of traditional scheduling that does not consider the impact of heat dissipation load on the total energy efficiency of nodes, so that the scheduling decision is more in line with the actual physical energy efficiency operation law.

[0006] 2. Addressing the issue of closed-loop coordination between scheduling strategies and real-time node status: To address the problem that native scheduling configurations cannot adapt to dynamic changes in node power consumption in real time, a custom dynamic update mechanism for ClassAd energy efficiency attributes was implemented, and the global RANK matching decision logic was reconstructed. This established a real-time feedback loop between the underlying node power consumption metrics and the top-level scheduling logic. This mechanism can drive a dynamic, energy-efficiency-oriented balanced distribution of job loads across a heterogeneous resource pool based on real-time energy efficiency feedback from nodes. While reducing overall cluster energy consumption, it effectively mitigates the risk of hardware degradation caused by localized heat buildup and improves system stability.

[0007] The technical solution of the present invention is as follows: An energy efficiency-aware scheduling method based on HTCondor for fused heat dissipation loss, comprising the following steps: S1. Construct a multidimensional power consumption decomposition model for nodes and perform fine decomposition of dynamic power consumption; S2. Introduce a multi-dimensional energy efficiency evaluation system with nonlinear heat dissipation penalty, that is, to measure the computing power output of a node's unit dynamic total power consumption, construct an energy efficiency evaluation model coupled with heat dissipation cost. S3 is compatible with native non-intrusive data acquisition / sensing mechanisms. In order to obtain input data that can truly reflect the energy efficiency status of nodes and support energy efficiency model calculations, it adopts a combination of standardized tools and native HTCondor attributes to achieve unified data acquisition of heterogeneous nodes. S4, ClassAd attribute extension and anti-jitter control, through timed sampling + dynamic threshold anti-jitter, achieve real-time and stable updates of the node's comprehensive energy efficiency score; S5. Global scheduling optimization is performed based on RANK priority reconstruction. At the HTCondor central management node, a global scheduling priority matching strategy is constructed with the node's comprehensive energy efficiency score as the core evaluation dimension. S6, dynamic closed-loop collaboration throughout the entire process.

[0008] According to a preferred embodiment of the present invention, in step S1, the model building process is as follows: First, a node total power consumption decomposition model is constructed to achieve a refined characterization of the sources of node energy consumption and the real-time total power consumption of the node. Represented as:

[0009] in, This represents the static standby baseline power consumption after the node is powered on and ready. At this time, the node is idle and has no business operations, and it is relatively stable under specific hardware conditions. This represents the node's dynamic total power consumption, which fluctuates dynamically with changes in load. Based on the actual operating conditions of the system, it can be further decomposed into: ; in, To calculate power consumption and CPU utilization It exhibits a non-linear positive correlation, reflecting the magnitude of the load; The heat dissipation power is generated by the operation of the cooling fan and is approximately proportional to the cube of the fan speed. The power consumption of hardware such as memory and hard drive accounts for a relatively low percentage and is relatively stable.

[0010] As can be seen from the above detailed breakdown of dynamic power consumption, computing power consumption and heat dissipation power consumption are the core components and sensitive fluctuation items of node dynamic power consumption, and are also the main sources of total energy consumption of computing clusters. In the process of job scheduling and distribution, identifying and prioritizing the execution of jobs by nodes with lower computing power consumption and heat dissipation power consumption is the key to achieving energy saving and consumption reduction and improving overall energy efficiency of computing clusters.

[0011] According to a preferred embodiment of the present invention, in step S2, specifically: S21. Establish a basic energy efficiency ratio; The basic computing power contribution is quantified by using node floating-point operation benchmarks and real-time CPU utilization to construct a basic energy efficiency ratio. : ; in, Indicates the node's floating-point operation capability (unit: kilo-float operations / second); This indicates the current CPU utilization of the node, with a value ranging from 0 to 1; This represents the base energy efficiency bias (ranging from 0.05 to 0.2), used to prevent abnormal fluctuations in scores under extremely low load scenarios, ensuring the smoothness and computational stability of the energy efficiency score. The denominator... The total power consumption of a node is the difference between its total power consumption and the static baseline power consumption, which is the node's dynamic total power consumption. S22. Construct a heat dissipation loss penalty operator; Based on the energy consumption characteristics of fluid dynamics, fan power consumption increases cubically with increasing speed. An exponential heat dissipation loss penalty operator is constructed to simulate the physical heat dissipation cost. : ; in, Indicates the fan duty cycle (percentage 0~100); A weight (ranging from 0.1 to 0.9) is assigned to adjust heat dissipation losses to accommodate differences in heat dissipation efficiency across different device models. A larger weight results in a more stringent energy efficiency penalty for nodes with high fan duty cycles. This operator characterizes the negative impact of nonlinear heat dissipation losses on the overall energy efficiency of nodes by imposing an "exponential penalty" on nodes with high fan duty cycles.

[0012] S23, Node comprehensive energy efficiency score; The node's overall energy efficiency score is obtained by multiplying the node's basic energy efficiency ratio by the heat dissipation loss penalty operator. This achieves deep coupling between computing power output and physical heat dissipation costs, resulting in a high score. The higher the score, the higher the effective computing power output per unit of energy consumption and the lower the proportion of ineffective heat dissipation loss, indicating a high-efficiency node. Conversely, the lower the score, the lower the node is due to hardware generation gaps, heat accumulation, or high fan load operation, resulting in a low computing power-to-power ratio and large heat dissipation redundancy, indicating a low-efficiency node.

[0013] According to a further preferred embodiment of the present invention, in step S23, The calculation formula is: ,Right now .

[0014] According to a preferred embodiment of the present invention, in step S3, specifically: S31, Physical Layer: Use IPMItool to obtain the node's real-time total power consumption through the standard IPMI interface. Duty cycle of fan speed As an industry-standard tool, IPMItool is compatible with server hardware from mainstream brands such as HPE, DELL, and Inspur, requiring no additional deployment dependencies, and provides underlying physical data support for power consumption breakdown and heat dissipation calculation. S32, Load Layer: Real-time collection of node CPU utilization via system-level monitoring tools (such as top or the underlying / proc statistics interface). It also reuses the native KFlops attribute of the HTCondor node as the node's floating-point operation capability. No additional data acquisition module needs to be developed, ensuring compatibility with the native scheduling framework and avoiding data redundancy and inconsistency issues. It leverages CPU utilization... With node floating-point arithmetic capabilities The actual computing power output of the product quantization node under the current load level.

[0015] According to a preferred embodiment of the present invention, in step S4, specifically: S41, Timed Sampling and ClassAd Injection; Lightweight sampling scripts are deployed on each computing node. Through the native STARTD_CRON mechanism of HTCondor, the scripts are executed automatically at a period of 5s to 30s to complete the collection of dynamic running data (real-time power consumption, load, fan duty cycle, etc. of the node) and the calculation of the node's comprehensive energy efficiency score. The final calculation results are injected into the node's custom ClassAd attribute: EnergyEfficiencyScore in real time to expose the node's real-time energy efficiency profile to the central scheduler. S42, Dynamic power consumption threshold anti-shake control; To avoid frequent abrupt changes in energy efficiency scores caused by instantaneous power fluctuations and sampling jitter in low-power states, a dynamic power consumption change threshold is introduced for smooth control. This threshold is applied only when the node's dynamic power consumption changes... When the preset threshold is exceeded, the node energy efficiency score is updated; otherwise, the historical score of the previous cycle remains unchanged. The threshold can be fine-tuned according to the cluster hardware characteristics, ranging from 5W to 20W. This can effectively avoid frequent switching of scheduling strategies caused by instantaneous sampling noise in low-power scenarios, prevent system scheduling oscillations, and improve the stability of cluster operation.

[0016] According to a preferred embodiment of the present invention, in step S5, specifically: The global priority expression of the central scheduler is reconstructed, and a node comprehensive energy efficiency score (EnergyEfficiencyScore) represented by a custom ClassAd attribute is introduced to reconstruct and optimize the resource allocation logic, thereby achieving intelligent scheduling that prioritizes node energy efficiency. NEGOTIATOR_PRE_JOB_RANK= (EnergyEfficiencyScore× )+((RemoteOwner=?= UNDEFINED)× -SlotID; Among them, NEGOTIATOR_PRE_JOB_RANK is a global preprocessing sorting expression used by the HTCondor central scheduler during the resource matching phase. The higher its calculated score, the higher the priority of the node being selected. and These are weighting coefficients, determined by setting orders of magnitude differences (e.g., ...). =1000, =100), ensuring that the node's overall energy efficiency score occupies the primary dominant position in multi-objective decision-making; RemoteOwner is an attribute of HTCondor used to represent the node's occupancy status, combined with the =?= matching operator, to determine whether the node is in the UNDEFINED idle state; SlotID is the resource slot number inside the node; The overall energy efficiency score of nodes is used as the primary basis for scheduling decisions, while taking into account the original node idle status and resource slot number constraints. This enables the central scheduler to automatically and preferentially allocate jobs to nodes / slots with higher energy efficiency scores, lower heat dissipation losses, and better idle status.

[0017] According to a preferred embodiment of the present invention, in step S6, specifically: Through a timed sampling and update mechanism, the EnergyEfficiencyScore value of the node's custom energy efficiency attribute is dynamically refreshed, and the RANK node matching priority strategy value is adjusted in real time. This process realizes the direct feedback of the node's physical operating characteristics (load status, power consumption changes, heat dissipation efficiency) to the top-level scheduling decision. At the same time, the updated scheduling priority acts in real time and guides the resource allocation of new jobs, realizing the dynamic closed-loop collaboration of the entire process of node operating condition perception, energy efficiency score calculation, RANK weight update, job scheduling execution, and node status feedback.

[0018] An energy efficiency-aware scheduling system based on HTCondor with integrated heat dissipation loss includes: The building module is used to construct a multidimensional power consumption decomposition model for nodes and to perform fine decomposition of dynamic power consumption; The evaluation module is used to construct a multi-dimensional energy efficiency evaluation system that incorporates nonlinear heat dissipation penalties. That is, to measure the computing power output of a node's unit dynamic total power consumption, an energy efficiency evaluation model coupled with heat dissipation costs is constructed. The acquisition module is designed to be compatible with native non-intrusive data acquisition / sensing mechanisms. To obtain input data that accurately reflects the energy efficiency status of nodes and support energy efficiency model calculations, it adopts a combination of standardized tools and native HTCondor attributes to achieve unified data acquisition from heterogeneous nodes. The update module is used to add custom ClassAd energy efficiency attributes and anti-jitter control. It achieves real-time and stable updates of the node's comprehensive energy efficiency score through timed sampling and dynamic threshold anti-jitter. The matching module optimizes global scheduling based on RANK priority reconstruction. At the HTCondor central management node, a global scheduling priority matching strategy is constructed with the node's comprehensive energy efficiency score as the core evaluation dimension. The collaboration module is used for dynamic closed-loop collaboration throughout the entire process.

[0019] The beneficial effects of this invention are as follows: 1. Constructing a multi-dimensional energy efficiency coupling model to improve computing power energy efficiency ratio: This invention constructs a multi-dimensional node energy efficiency evaluation model by quantifying the correlation between unit computing power power consumption and dynamic heat dissipation loss, solving the technical deficiency of traditional scheduling that relies solely on nominal hardware attributes while ignoring heat dissipation costs. This model can quantify the energy efficiency differences between heterogeneous nodes, guiding jobs to prioritize high-energy-efficiency nodes with lower unit computing power power consumption and less heat dissipation loss, effectively avoiding low-energy-efficiency nodes with high heat dissipation redundancy. By reducing the redundant power consumption of low-energy-efficiency nodes, energy saving and consumption reduction are achieved, effectively improving the overall energy efficiency ratio and energy utilization efficiency of the cluster.

[0020] 2. Achieving closed-loop collaboration between scheduling strategies and real-time node status to enhance dynamic adaptability: This invention achieves accurate perception of the dynamic energy efficiency characteristics of cluster nodes by collecting real-time power consumption and operating parameters of computing nodes, combined with the HTCondor dynamic attribute injection mechanism. By deeply integrating real-time energy efficiency scores into the RANK matching strategy, it ensures that jobs always prioritize matching high-energy-efficiency nodes in the current cluster state. Simultaneously, changes in the node's operating environment (such as heat accumulation and power fluctuations) update its own ClassAd attribute in real time, which in turn affects the next round of scheduling decisions, forming an adaptive closed-loop collaborative mechanism of "perception-decision-feedback-optimization." This effectively solves the problem that traditional resource-oriented scheduling models cannot respond to the dynamic evolution of the operating environment, improving the scheduling agility of heterogeneous clusters under complex operating conditions.

[0021] 3. Low implementation cost and strong compatibility: This invention does not require modification of the HTCondor native architecture. Energy efficiency-aware scheduling upgrades can be achieved by adding a custom ClassAd attribute and reconstructing the global RANK matching strategy. This non-intrusive technical solution not only ensures the operational stability of the original scheduling system but also achieves smooth integration and rapid deployment of functions. The solution possesses excellent cross-platform compatibility, seamlessly adapting to homogeneous clusters and complex heterogeneous environments with mixed new and old hardware and multiple brands of hardware, making it highly valuable for engineering implementation.

[0022] 4. Mitigating Localized Heat Accumulation and Hardware Fatigue, Optimizing Cluster Operations and Maintenance: Addressing the energy efficiency imbalance caused by differences in the age, brand, and configuration of equipment in heterogeneous clusters, this invention utilizes real-time energy efficiency sensing and a heat dissipation penalty mechanism to avoid the risk of localized heat accumulation caused by excessive concentration of operations on low-energy-efficiency nodes. This dynamic energy-efficiency-oriented scheduling effectively reduces the risk of equipment throttling or downtime due to high temperatures, slows down hardware aging and wear, improves operational stability while reducing long-term data center operation and maintenance costs, and has significant implications for green and sustainable development. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0025] Example 1: This embodiment provides an energy efficiency-aware scheduling method based on HTCondor for fused heat dissipation loss, the steps of which are as follows: I. Node-level Physical Status Awareness and Automated Data Acquisition: Lightweight data acquisition scripts are written and deployed on each compute node. Utilizing HTCondor's STARTD_CRON mechanism, the data acquisition scripts are registered as periodic tasks directly managed by the scheduling framework. The scripts primarily acquire real-time operating parameters of the nodes periodically through standard hardware interfaces (such as IPMI), system-level monitoring interfaces, and HTCondor native attributes, including the node's real-time total power consumption. Fan duty cycle CPU utilization and node floating-point arithmetic capabilities This provides underlying physical data support for subsequent node energy efficiency assessments;

[0026] II. Node-level refined power consumption breakdown and energy efficiency score calculation; After the underlying data collection is completed, numerical calculations are performed based on the preset energy efficiency evaluation model: (1) Calculation of the basic energy efficiency ratio of nodes: ; in, Indicates the node's floating-point operation capability (unit: kilo-float operations / second); This indicates the current CPU utilization of the node, with a value ranging from 0 to 1; This represents the base energy efficiency bias (ranging from 0.05 to 0.2), used to prevent abnormal fluctuations in scores under extremely low load scenarios, ensuring the smoothness and computational stability of the energy efficiency score. The denominator... The total power consumption of a node is the difference between its total power consumption and the static baseline power consumption, which is the node's dynamic total power consumption. Its physical significance lies in quantifying the effective computing power output per unit of dynamic power consumption of a node under the current load level, and introducing a basic energy efficiency bias coefficient. To avoid numerical anomalies under extremely low loads and improve model robustness.

[0027] (2) Calculation of heat dissipation penalty operator: ; in, Indicates the fan duty cycle (percentage 0~100); A weight (ranging from 0.1 to 0.9) is assigned to adjust heat dissipation losses to accommodate differences in heat dissipation efficiency across different device models. A larger weight results in a more stringent energy efficiency penalty for nodes with high fan duty cycles. This operator characterizes the negative impact of nonlinear heat dissipation losses on the overall energy efficiency of nodes by imposing an "exponential penalty" on nodes with high fan duty cycles.

[0028] The core logic of introducing a third-order heat dissipation penalty operator lies in using the cube of the fan duty cycle as a penalty variable to construct a nonlinear penalty model that conforms to the laws of fluid dynamics. This is used to fine-tune the sensitivity of fan energy consumption characteristics for different physical architectures, making the algorithm universal. For example, high-density 1U rack servers, due to their limited internal space and high fan speeds, The value should be greater than that of a standard 2U or 4U rack server.

[0029] (3) Calculation of node comprehensive energy efficiency score: This formula integrates the computational output efficiency of nodes with the environmental heat dissipation cost, breaking away from the traditional energy-saving scheduling that only considers the single dimension of CPU utilization or power consumption, and constructing a multi-dimensional comprehensive evaluation system that takes into account computing power, energy consumption and environmental costs.

[0030] III. Logical Attribute Mapping and Debouncing Control: A custom ClassAd energy efficiency attribute, EnergyEfficiencyScore, is created within each compute node. The node's overall energy efficiency score is mapped into the scheduling system layer, making it globally observable. Simultaneously, to prevent score fluctuations and subsequent scheduling decision oscillations caused by instantaneous sampling noise, a dynamic power consumption threshold is set for each node. Power consumption is only applied when... Only when the preset threshold is exceeded is it considered a valid state evolution, thereby triggering an update of the EnergyEfficiencyScore value of the node's ClassAd attribute.

[0031] IV. Top-level global decision-making matching stage; At the central management node, the central scheduler (Negotiator) retrieves the latest ClassAd attribute of all idle computing nodes / slots in real time during each job distribution cycle, and focuses on extracting the EnergyEfficiencyScore dynamically generated by each node in the above stage. Based on the reconstructed RANK priority matching strategy, the system uses this energy efficiency score as the core weight factor to automatically calculate the matching priority of each node / slot. Finally, based on the calculation results, the jobs to be processed are preferentially distributed to nodes with higher energy efficiency scores.

[0032] V. Dynamic closed-loop feedback and continuous regulation; The job starts running on the matched target node. As the node's computational load increases, the node's dynamic total power consumption increases. and cooling fan duty cycle As the number of nodes gradually increases, the data collection scripts deployed on the nodes capture the above changes, recalculate the node's overall energy efficiency score, and synchronize it to the resource pool in real time through the ClassAd attribute. In the subsequent scheduling cycle, the central scheduler re-prioritizes the nodes based on the reconstructed priority matching strategy. Newly submitted job tasks are automatically guided to nodes with higher heat dissipation redundancy and better energy efficiency scores, ultimately forming a dynamic closed loop of the entire process: "node operating condition perception - energy efficiency score calculation - RANK weight update - job scheduling execution - node status feedback".

Claims

1. An energy efficiency-aware scheduling method based on HTCondor for fused heat dissipation loss, characterized in that, The steps are as follows: S1. Construct a multi-dimensional power consumption decomposition model for nodes to deconstruct dynamic power consumption; S2. Introduce a multi-dimensional energy efficiency evaluation system with nonlinear heat dissipation penalty, that is, to measure the computing power output of a node's unit dynamic total power consumption, construct an energy efficiency evaluation model coupled with heat dissipation cost. S3 is compatible with native non-intrusive data acquisition / sensing mechanisms. In order to obtain input data that can truly reflect the energy efficiency status of nodes and support energy efficiency model calculations, it adopts a combination of standardized tools and native HTCondor attributes to achieve unified data acquisition of heterogeneous nodes. S4, ClassAd attribute extension and anti-jitter control, through timed sampling + dynamic threshold anti-jitter, achieve real-time and stable updates of the node's comprehensive energy efficiency score; S5. Global scheduling optimization is performed based on RANK priority reconstruction. At the HTCondor central management node, a global scheduling priority matching strategy is constructed with the node's comprehensive energy efficiency score as the core evaluation dimension. S6, dynamic closed-loop collaboration throughout the entire process.

2. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 1, characterized in that, In step S1, the model building process is as follows: First, a node total power consumption decomposition model is constructed to characterize the sources of energy consumption and the node's real-time total power consumption. Represented as: ； in, This represents the static standby reference power consumption after the node is powered on and ready. The dynamic total power consumption of a node is represented, and it can be further decomposed according to the actual operating conditions as follows: ； in, To calculate power consumption; For heat dissipation power consumption; It refers to hardware power consumption.

3. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 2, characterized in that, In step S2, specifically: S21. Establish a basic energy efficiency ratio; The basic computing power contribution is quantified by using node floating-point operation benchmarks and real-time CPU utilization to construct a basic energy efficiency ratio. : ； in, Indicates the node's floating-point arithmetic capability; This indicates the current CPU utilization of the node; Indicates the basic energy efficiency bias, denominator The total power consumption of a node is the difference between its total power consumption and the static baseline power consumption, which is the node's dynamic total power consumption. S22. Construct a heat dissipation loss penalty operator; Based on the energy consumption characteristics of fluid dynamics, fan power consumption increases cubically with increasing speed. An exponential heat dissipation loss penalty operator is constructed to simulate the physical heat dissipation cost. : ； in, Indicates the fan duty cycle; The weights for heat dissipation loss are adjusted to accommodate the differences in heat dissipation efficiency among different models. The larger the weight, the more severe the energy efficiency penalty for nodes with high fan duty cycles. S23, Node comprehensive energy efficiency score; The node's overall energy efficiency score is obtained by multiplying the node's basic energy efficiency ratio by the heat dissipation loss penalty operator. This achieves deep coupling between computing power output and physical heat dissipation costs, resulting in a high score. A higher score indicates a higher effective computing power output per unit of energy consumption and a lower proportion of ineffective heat dissipation losses, classifying it as a high-efficiency node; conversely, a lower score indicates a low computing power-to-power ratio and a large heat dissipation redundancy, classifying it as a low-efficiency node.

4. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 3, characterized in that, In step S23, The calculation formula is: ,Right now .

5. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 4, characterized in that, In step S3, specifically: S31, Physical Layer: Use IPMItool to obtain the node's real-time total power consumption through the standard IPMI interface. Duty cycle of fan speed ; S32, Load Layer: Real-time monitoring of node CPU utilization is achieved through system-level monitoring tools. It also reuses the native KFlops attribute of the HTCondor node as the node's floating-point operation capability. Utilize CPU utilization With node floating-point arithmetic capabilities The actual computing power output of the product quantization node under the current load level.

6. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 5, characterized in that, In step S4, specifically: S41, Timed Sampling and ClassAd Injection; Lightweight sampling scripts are deployed on each computing node. Through the native STARTD_CRON mechanism of HTCondor, the scripts are automatically executed every 5 to 30 seconds to complete the collection of dynamic running data and the calculation of the node's comprehensive energy efficiency score. The final calculation results are injected into the node's custom ClassAd attribute: EnergyEfficiencyScore in real time to expose the node's real-time energy efficiency profile to the central scheduler. S42, Dynamic power consumption threshold anti-shake control; To avoid frequent abrupt changes in energy efficiency scores caused by instantaneous power fluctuations and sampling jitter in low-power node states, a dynamic power consumption change threshold is introduced for smooth control. This threshold is applied only when the node's dynamic power consumption changes... The node's energy efficiency score is updated when the preset threshold is exceeded; otherwise, the historical score from the previous cycle remains unchanged.

7. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 6, characterized in that, In step S5, specifically: The global priority expression of the central scheduler is reconstructed, and a node comprehensive energy efficiency score (EnergyEfficiencyScore) represented by a custom ClassAd attribute is introduced to reconstruct and optimize the resource allocation logic, thereby achieving intelligent scheduling that prioritizes node energy efficiency. NEGOTIATOR_PRE_JOB_RANK= (EnergyEfficiencyScore× )+((RemoteOwner=?= UNDEFINED)× )–SlotID; Among them, NEGOTIATOR_PRE_JOB_RANK is a global preprocessing sorting expression used by the HTCondor central scheduler during the resource matching phase. The higher its calculated score, the higher the priority of the node being selected. and This is a weighting coefficient, which, by setting orders of magnitude differences, ensures that the node's overall energy efficiency score occupies the primary dominant position in multi-objective decision-making; RemoteOwner is an attribute of HTCondor used to represent the node's occupancy status, and, combined with the =?= matching operator, is used to determine whether the node is in an UNDEFINED idle state; SlotID is the resource slot number inside the node; The overall energy efficiency score of nodes is used as the primary basis for scheduling decisions, while taking into account the original node idle status and resource slot number constraints, so that the central scheduler prioritizes the allocation of jobs to nodes / slots with higher energy efficiency scores, lower heat dissipation losses, and better idle status.

8. The energy efficiency-aware scheduling method for fused heat dissipation loss based on HTCondor as described in claim 7, characterized in that, In step S6, specifically: Through a timed sampling and update mechanism, the node's custom energy efficiency attribute, EnergyEfficiencyScore, is dynamically refreshed, and the RANK node matching priority strategy value is adjusted in real time. This process enables direct feedback of the node's physical operating characteristics to the top-level scheduling decision. At the same time, the updated scheduling priority acts in real time and guides the resource allocation of new jobs, realizing dynamic closed-loop collaboration of the entire process from node condition perception to energy efficiency score calculation, RANK weight update, job scheduling execution, and node status feedback.

9. An energy efficiency sensing and scheduling system based on HTCondor that integrates heat dissipation loss, characterized in that, include: The building module is used to construct a multidimensional power consumption decomposition model for nodes and deconstruct dynamic power consumption; The evaluation module is used to construct a multi-dimensional energy efficiency evaluation system that incorporates nonlinear heat dissipation penalties. That is, to measure the computing power output of a node's unit dynamic total power consumption, an energy efficiency evaluation model coupled with heat dissipation costs is constructed. The acquisition module is designed to be compatible with native non-intrusive data acquisition / sensing mechanisms. To obtain input data that accurately reflects the energy efficiency status of nodes and support energy efficiency model calculations, it adopts a combination of standardized tools and native HTCondor attributes to achieve unified data acquisition from heterogeneous nodes. The update module is used to add ClassAd energy efficiency attributes and anti-jitter control. It achieves real-time and stable updates of the node's comprehensive energy efficiency score through timed sampling and dynamic threshold anti-jitter. The matching module optimizes global scheduling based on RANK priority reconstruction. At the HTCondor central management node, a global scheduling priority matching strategy is constructed with the node's comprehensive energy efficiency score as the core evaluation dimension. The collaboration module is used for dynamic closed-loop collaboration throughout the entire process.

Images