Computing node power consumption monitoring and power consumption prediction method

Abstract

The invention discloses a computing node power consumption monitoring and predicting method, and belongs to the technical field of computers. The problem that an existing power consumption prediction method is poor in accuracy is solved. According to the method, the dynamic operation states of the nodes, the hardware static configuration attributes and the coupling influence of the communication relation and power consumption between the nodes are comprehensively considered, a graph structure based on node communication is constructed, and a power consumption coupling path between the nodes is modeled and calculated, so that the dynamic energy consumption of local nodes can be captured; and cross-node resource contention, communication bottleneck and energy consumption conduction effect can be considered. Through a structure-guided graph attention mechanism, the model can realize information sharing and feature collaboration among a plurality of nodes. By introducing an energy efficiency attention fusion module, dynamic weighted modeling is performed on multi-channel input features, key factors having high influence on energy consumption fluctuation are automatically identified, and the accuracy and interpretability of the model in the aspects of power consumption prediction and optimization suggestion output are improved. The method can be applied to power consumption prediction of the computing node.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer, and particularly relates to a kind of computing node power consumption monitoring and power consumption prediction method. BACKGROUND

[0002] With the rapid development of cloud computing and Internet technology, the energy consumption caused by processing massive data has become a problem that cannot be ignored. High energy consumption leads to high cost. For example, the power consumption of green data centers is continuously rising. Among them, the electricity consumption of IT equipment accounts for 70% of the total power consumption of data centers, and the electricity cost of IT equipment alone exceeds 200 million yuan. As can be seen, the high energy consumption of equipment, the high cost of operation and the sustainable development of the problem of low utilization rate of energy consumption have become problems that the information and communication technology industry urgently needs to solve.

[0003] “Green” has been widely used to describe things with environmental friendly features, such as low carbon, energy saving, economic, sustainable, etc. Green computing, that is, from the design, manufacturing, transportation, use, maintenance, recycling and reuse of the whole life cycle process, uses computing resources in an economic and low-power consumption way (Sun Da-wei, Chang Gui- ran, Chen Dong, et al. Analysis, quantification, modeling and evaluation of green service level objectives in cloud computing environment [J]. Journal of Computer Science and Technology, 2013, 36(7): 17.). Green computing is put into practice, that is, green data center, whose goal is to achieve the sustainable development of data center and reduce the high pollution of IT equipment high power consumption to the environment. In the field of green computing, many domestic and foreign scholars have put forward their own solutions. Song Jie studies the optimal resource ratio of Map / Reduce task, and proposes an energy-efficient task scheduling and resource allocation algorithm (Song Jie, Liu Xuebing, Zhu Zhiliang, et al. An energy-efficient MapReduce resource ratio model [J]. Journal of Computer Science and Technology, 2015, 38(1): 15.). In the same year, Song Jie further proposed the data layout algorithm of MapReduce (Song Jie, Wang Zhi, Li Tiantian, et al. A data layout algorithm for optimizing the energy consumption of MapReduce system [J]. Journal of Software, 2015, 26(8): 20.). The next year, Song Jie proposed the new concept of algorithm energy consumption complexity, and redefined the software energy consumption evaluation model (Song Jie, Sun Zongzhe, Li Tiantian, et al. Research progress of code-oriented software energy consumption optimization [J]. Journal of Computer Science and Technology, 2016, 39(11): 21.). In addition, Liao Bin designs an energy-saving algorithm to improve the energy-saving of HDFS from the aspects of data block storage structure and storage strategy (Liao Bin, Yu Jing, Zhang Tao, et al. Energy-saving algorithm based on distributed file system HDFS [J]. Journal of Computer Science and Technology, 2013, 36(5): 18.). Yang Ting proposes a differential storage energy-saving optimization algorithm to realize the opening of fewer DataNodes and effectively reduce the power consumption of HDFS cluster (Yang Ting, Wang Meng, Zhang Yajian, et al. Cloud computing data center HDFS differential storage energy-saving optimization algorithm [J]. Journal of Computer Science and Technology, 2019, 42(04): 47-61.).Guo Chaopeng summarized the basic ideas of neural network energy consumption estimation methods and neural network lightweight methods (Guo Chaopeng, Wang Xinxin, Zhong Zhaojin, et al. Research Progress of Energy Consumption Optimization of Neural Network Lightweight Method [J]. Journal of Computer Research and Development, 2023, 46(1): 85-102.); Xiao Peng proposed a virtual machine scheduling algorithm based on energy consumption ratio model for mixed load scheduling problem (Xiao Peng, Liu Dongbo, Qu Xilong. Virtual Machine Scheduling Algorithm Based on Energy Consumption Ratio Model in Cloud Computing [J]. Electronics Letters, 2015, 43(2): 7.); Wang Haifeng introduced model predictive control strategy from the perspective of cluster global optimization to optimize the energy consumption of GPU cluster (Wang Haifeng, Cao Yunpeng. Research on Energy Consumption Optimization Control Model of GPU Cluster [J]. Electronics Letters, 2015, 43(10): 7.); Luo Liang proposed a server energy consumption model suitable for cloud computing data center (Luo Liang, Wu Wenjun, Zhang Fei. Energy Consumption Modeling Method for Cloud Computing Data Center. . Journal of Software, 2014, (7): 1371-87.); Wang Zhaoguo proposed a machine learning-oriented model for machine learning, a computationally intensive application (Wang Zhaoguo, Yi Han, Zhang Weihua. Data Center Energy Consumption Optimization Method Based on Machine Learning Characteristics [J]. Journal of Software, 2014, 25(7): 16.); Virtualization technology has been developed and widely applied, Ye Kejiang pointed out the difficulties of energy consumption monitoring and measurement of virtualization cloud computing platform (Ye Kejiang, Wu Zhaohui, Jiang Xiaohong, et al. Energy Consumption Management of Virtualization Cloud Computing Platform [J]. Journal of Computer Research and Development, 2012, 35(6): 24.); At the same time, Donghyun Gouk based on container technology realized DockerSSD, by offloading part of the computing task, running independently in the ISP-container in the SSD, to reduce the energy consumption of the host (GOUK D, KWON M, BAE H, et al. DockerSSD: Containerized In-Storage Processing and Hardware Acceleration for Computational SSDs; proceedings of the 2024 IEEE International Symposium on High-Performance Computer Architecture (HPCA), F, 2024 [C]. IEEE.).

[0004] Currently in the field of green computing, the energy consumption measurement and management method of data center mainly focuses on hardware, software, model, virtualization technology or simulation technology (Lin Weiwei, Wu Wentai. Energy consumption measurement and management method for cloud computing environment [J]. Journal of software, 2016, (4): 16.), however, the data reliability of virtualization technology and simulation technology is not high, and the type of hardware environment of simulation device is limited, the data accuracy of software-based direct measurement method (using DAQ system) is high, but the deployment is not easy, and the migration is low.

[0005] On the other hand, the existing power consumption modeling method mainly focuses on single or small feature dimension, and the common method includes modeling based on system usage rate (such as CPU, memory, storage, network) or based on underlying performance counter (such as instruction number, cache hit, page table access) (Luo Liang, Wu Wentai, Zhang Fei. Energy consumption modeling method for cloud computing data center . Journal of software, 2014, (7): 1371-87.). However, the existing method does not consider the influence of the dynamic running state of the node, the static configuration attribute of the hardware (such as CPU core number, memory capacity) and the communication relationship between nodes on the coupling effect of power consumption, therefore, the accuracy of the existing power consumption prediction method is still poor, and it is necessary to propose a new prediction method to solve the above problems. SUMMARY

[0006] The purpose of the application is to solve the problem of poor accuracy of the existing power consumption prediction method, and a computing node power consumption monitoring and power consumption prediction method is proposed.

[0007] The technical scheme adopted by the application to solve the above technical problems is: a computing node power consumption monitoring and power consumption prediction method, the method specifically comprises the following steps:

[0008] Step one, obtaining the static configuration parameters of the computing node i, using the power meter to collect the power consumption of the computing node i in the process of executing the work load, and collecting the dynamic state parameters of the computing node i in the process of executing the work load through Linux instruction;

[0009] Step two, aligning the time stamp of the dynamic state parameters of the computing node i with the time stamp of the power consumption data of the computing node i, to obtain the synchronous time sequence data of the dynamic state parameters and the power consumption of the computing node i;

[0010] Among them, the time sequence data of the power consumption of the computing node i is , is a real number, is the active power of the th sampling point, is the sampling length;

[0011] The time series data of the dynamic state parameter of the computing node i is , , is the dynamic state parameter sequence of the i-th sampling point, is the CPU utilization of the i-th sampling point, is the memory occupancy of the i-th sampling point, is the network input rate of the i-th sampling point, is the network output rate of the i-th sampling point.

[0012] Step three, normalize and difference process the time series data to obtain the differential sequence data of power consumption; normalize and difference process the data of each dimension in the time series data to obtain the differential sequence data of CPU utilization, the differential sequence data of memory occupancy, the differential sequence data of network input rate and the differential sequence data of network output rate;

[0013] Align the differential sequence data of power consumption, CPU utilization, memory occupancy, network input rate and network output rate of the computing node i to form an input tensor matrix of the computing node i;

[0014] Step four, construct a static attribute vector according to the static configuration parameter of the computing node i, and map the static attribute vector into a dense vector;

[0015] Obtain the fusion feature of the computing node i according to the input tensor matrix and the dense vector of the computing node i;

[0016] Step five, take the fusion feature of the computing node i as the input of the power consumption prediction model deployed on the computing node i, and output the power consumption prediction result of the computing node i through the power consumption prediction model.

[0017] Further, the static configuration parameter specifically includes the number of CPU cores, the size of memory, the type of hard disk and the node role.

[0018] Further, the dynamic state parameter specifically includes CPU utilization, memory occupancy, network input rate and network output rate.

[0019] Further, the normalization and difference process of the time series data to obtain the differential sequence data of power consumption; specifically,

[0020] ​​​​​

[0021] wherein, is the mean of all data in the power consumption time series data ;

[0022] is the standard deviation of all data in the power consumption time series data ;

[0023] is the corresponding normalized value in the power consumption time series data ;

[0024] then the value of the differential sequence data of the power consumption is:

[0025] .

[0026] Further, the static attribute vector of the computing node i is constructed according to the static configuration parameters of the computing node i , specifically:

[0027]

[0028] wherein, c is the number of CPU cores of the computing node i;

[0029] m is the memory size of the computing node i;

[0030] d is the hard disk type of the computing node i;

[0031] role is the role of the computing node i;

[0032] is One-Hot encoding;

[0033] is the static attribute vector;

[0034] is mapped into a dense vector through embedding:

[0035]

[0036] wherein, represents the dense vector;

[0037] represents a mapping process.

[0038] Further, the fusion feature of the computing node i is obtained according to the input tensor matrix and the dense vector of the computing node i, specifically:

[0039] ​​Step four one:

[0040]

[0041] wherein, represents the feature corresponding to the th sampling point in the input tensor matrix of the computing node i;

[0042] represents the embedding encoder;

[0043] represents the position encoding of the th sampling point;

[0044] represents the intermediate variable;

[0045] Step four two: according to obtain the fusion feature of the computing node i at the th sampling point:

[0046]

[0047] wherein, represents the feature of the th channel in

[0048] represents the total number of channels;

[0049] represents the fusion feature of the computing node i at the th sampling point;

[0050] represents the weight of the th channel of the th sampling point;

[0051] Step four three: the features corresponding to each sampling point in the input tensor matrix of the computing node i are processed by the method of step four one and step four two, to obtain the fusion feature of the computing node i at each sampling point, and the fusion features of various sampling points are used to form the fusion feature of the computing node i.

[0052] Further, the calculation method of the weight of the th channel of the th sampling point is:

[0053] the feature of the th channel is mapped to a scalar score by MLP:

[0054] ​

[0055] Again Normalized by Softmax weight mechanism as weight :

[0056]

[0057] Wherein, And .

[0058] Further, the specific process of step five is:

[0059] (1) When the number of computing nodes in the cluster that need to predict power consumption is 1, the power consumption prediction model deployed on the computing node is a standard Transformer;

[0060] (2) When the number of computing nodes in the cluster that need to predict power consumption is greater than 1, the power consumption prediction model deployed on each computing node is an improved Transformer;

[0061] The improved Transformer is specifically: the Transformer model after replacing the attention mechanism in the standard Transformer encoder with a graph attention mechanism;

[0062] The input of the improved Transformer is: the fusion features of the computing node where the improved Transformer is located and the adjacency matrix between all computing nodes in the cluster;

[0063] The calculation method of the adjacency matrix between all computing nodes in the cluster is method 1, method 2 or method 3;

[0064] Method 1: If the i-th node communicates with the l-th node, the element in the i-th row and the l-th column of the adjacency matrix is =1, otherwise the element in the i-th row and the l-th column of the adjacency matrix is =0, that is, the adjacency matrix is a binary adjacency matrix composed of 0 and 1;

[0065] Method 2: The element in the i-th row and the l-th column of the adjacency matrix is , = , Indicates the bandwidth inverse ratio between the i-th node and the l-th node;

[0066] Method 3: The element in the i-th row and the l-th column of the adjacency matrix is :

[0067]

[0068] Wherein, This represents the communication frequency between the i-th node and the l-th node;

[0069] This indicates the total number of nodes in the cluster.

[0070] Furthermore, the specific process of step six is ​​as follows:

[0071] Based on the output of the last attention layer of the power consumption prediction model, the importance ranking of each dynamic state parameter is obtained, and optimization suggestions are generated based on the importance ranking results.

[0072] Furthermore, the loss function used during the training of the power consumption prediction model... for:

[0073]

[0074] in, Represents the cross-entropy loss function;

[0075] This refers to Huber Loss;

[0076] and It's a hyperparameter.

[0077] The beneficial effects of this invention are:

[0078] This invention comprehensively considers the dynamic operating status of nodes, static hardware configuration attributes, and the coupling effect of inter-node communication relationships and power consumption. It also introduces an energy efficiency attention fusion module to perform dynamic weighted modeling on multi-channel input features, automatically identify key factors that have a high influence on energy consumption fluctuations, thereby improving the accuracy and interpretability of the model in power consumption prediction and optimization suggestion output.

[0079] This invention is applicable not only to single-node cluster environments but also to multi-node cluster scenarios. In multi-node modeling, the system further constructs a graph structure based on node communication, modeling and computing the power consumption coupling paths between nodes. This allows it to not only capture the energy consumption dynamics of local nodes but also consider cross-node resource contention, communication bottlenecks, and energy consumption transmission effects. Through a structure-guided graph attention mechanism, the model can achieve information sharing and feature collaboration among multiple nodes, effectively supporting the generation of optimization strategies and realizing cluster-level energy efficiency linkage optimization. This invention is a complete integrated system for data acquisition, prediction optimization, and visualization platforms, supporting a closed-loop cluster-level energy efficiency management system. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of computing node power consumption.

[0081] Figure 2 This is a flowchart of a computing node power consumption monitoring and prediction method according to the present invention;

[0082] Figure 3 This is a schematic diagram of the model's working process;

[0083] Figure 4 This is a flowchart of attention interpretation generation and suggestion mapping;

[0084] Figure 5 This is a diagram of the system task submission page;

[0085] Figure 6 This is a schematic diagram of the system power consumption monitoring page; Detailed Implementation

[0086] Specific Implementation Method 1: Combination Figure 2 This embodiment describes a method for monitoring and predicting power consumption of a computing node. The method specifically includes the following steps:

[0087] Step 1: Obtain the static configuration parameters of compute node i by executing the acquisition scripts of commands lscpu, free and lsblk, collect the power consumption of compute node i in real time during the execution of the workload using a power meter, and collect the dynamic status parameters of compute node i in real time during the execution of the workload using Linux commands.

[0088] Step 2: Align the timestamps of the dynamic state parameters of computing node i with the timestamps of the power consumption data of computing node i (since the sampling frequency of the power consumption data is less than the sampling frequency of the dynamic state parameters, the timestamps of the power consumption data are used as the alignment reference, and the dynamic state parameters corresponding to each timestamp of the power consumption data are retained) to obtain the synchronous time series data of the dynamic state parameters and power consumption of computing node i.

[0089] The time-series data of the power consumption of computing node i is as follows: , For real numbers, For the first Active power at each sampling point The sampling length;

[0090] The time series data of the dynamic state parameters of node i are: , , For the first The dynamic state parameter sequence of each sampling point For the first CPU utilization per sampling point For the first Memory usage per sampling point For the first Network input rate at each sampling point For the first Network output rate per sampling point;

[0091] Step 3: Process time series data Normalization and differencing are performed to obtain the power consumption differential sequence data; the time series data... The data for each dimension are normalized and differentially processed to obtain differential sequence data for CPU utilization, memory usage, network input rate, and network output rate.

[0092] Align the differential sequence data of power consumption, CPU utilization, memory usage, network input rate, and network output rate of computing node i to form the input tensor matrix of computing node i (the first row of the matrix is ​​the differential sequence data of power consumption, the second row is the differential sequence data of CPU utilization, the third row is the differential sequence data of memory usage, the fourth row is the differential sequence data of network input rate, and the fifth row is the differential sequence data of network output rate).

[0093] Step 4: Construct a static attribute vector based on the static configuration parameters of compute node i. and static attribute vector Mapped into a dense vector;

[0094] The fusion features of computing node i are obtained based on the input tensor matrix and dense vector of computing node i.

[0095] Step 5: Use the fusion features of computing node i as input to the power prediction model deployed on computing node i, and output the power prediction result of computing node i through the power prediction model.

[0096] This implementation uses a hardware-based direct measurement method for power consumption. This method is characterized by low deployment difficulty, high data reliability, and low node granularity. For example... Figure 1The diagram illustrates a platform for collecting power consumption and compute node resource information. All compute nodes are Linux systems. The cluster has two compute nodes. Assuming power consumption prediction modeling is needed for compute node B, node B is connected to a power meter via USB, which in turn connects to a power supply. The external power meter is an HP9800. The power meter collects real-time power consumption data and transmits it to the power prediction system embedded on the compute node. Node A only needs to be connected to a power supply to ensure the cluster operates normally. Power consumption prediction modeling can be performed on both compute nodes A and B. The difference between performing power consumption prediction modeling on one compute node and multiple compute nodes lies only in the subsequent prediction model input. The workloads are from the open-source big data benchmarking toolkit HiBench, such as Sort, WordCount, and TeraSort in Micro Benchmark-type workloads, K-means in machine learning workloads, and PageRank in graph computing workloads.

[0097] This invention presents a Transformer-based node power consumption prediction method that not only outputs predicted values ​​but also automatically provides optimization suggestions for the current state. It interprets the model's focus through attention weights, thereby providing decision support for system optimization. Finally, this implementation supports a visual interface to display the hardware and software status of the data center cluster, helping operations and maintenance personnel quickly grasp the cluster's condition.

[0098] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the static configuration parameters specifically include the number of CPU cores, memory size, hard disk type, and node role.

[0099] The other steps and parameters are the same as in Specific Implementation Method 1.

[0100] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that the dynamic state parameters specifically include CPU utilization, memory usage, network input rate, and network output rate.

[0101] Other steps and parameters are the same as in specific implementation method one or two.

[0102] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the time series data... Normalization and differencing are performed to obtain the power consumption differential sequence data; specifically:

[0103]

[0104] in, Power consumption time series data The mean of all data in the dataset;

[0105] Power consumption time series data The standard deviation of all data in the dataset;

[0106] Power consumption time series data In The corresponding normalized value;

[0107] Then the first in the differential sequence data of power consumption Value for:

[0108]

[0109] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0110] The method of this embodiment can process the data of each dimension in sequence D separately, that is, process the CPU utilization rate at each time step, process the memory usage rate at each time step, process the network input rate at each time step, and process the network output rate at each time step.

[0111] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that it involves constructing a static attribute vector based on the static configuration parameters of computing node i. Specifically:

[0112]

[0113] Where c is the number of CPU cores in computing node i;

[0114] m is the memory size of computing node i;

[0115] d represents the hard disk type of compute node i;

[0116] role is the role of compute node i (it can be a master node or a slave node; in the case of a single node, it defaults to the master node).

[0117] Encode for One-Hot;

[0118] It is a static attribute vector;

[0119] By embedding Mapping to dense vectors (i.e., the standard structure of mapping discrete symbols to dense vectors using an embedding layer):

[0120]

[0121] in, Represents a dense vector;

[0122] This indicates mapping processing.

[0123] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0124] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the step of obtaining the fusion features of computing node i based on the input tensor matrix and dense vector of computing node i is specifically as follows:

[0125] Step 41:

[0126]

[0127] in, This represents the first element in the input tensor matrix of computation node i. Features corresponding to each sampling point (including the t-th power consumption value, the t-th CPU utilization value, the t-th memory usage value, the t-th network input rate value, and the t-th network output rate value).

[0128] This indicates an embedded encoder (the present invention uses a GNN network).

[0129] Indicates the first The location code of each sampling point;

[0130] Indicates intermediate variables. d is a unified embedding dimension, which can be adjusted based on engineering experience or actual data;

[0131] Step 42: According to Obtain the computation node i at the th Fusion features of each sampling point:

[0132]

[0133] in, express Middle Characteristics of each channel;

[0134] Indicates the total number of channels;

[0135] Indicates that the computation node i is at the 1st epoch. Fusion features of each sampling point;

[0136] Indicates the first The sampling point of the first sampling point The weight of each channel;

[0137] Step 43: Using the methods in Step 41 and Step 42, process the features corresponding to each sampling point in the input tensor matrix of computing node i to obtain the fusion features of computing node i at each sampling point, and use the fusion features of each sampling point to form the fusion features of computing node i.

[0138] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0139] The process of this implementation is defined as an energy efficiency attention fusion mechanism.

[0140] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the first... The sampling point of the first sampling point The method for calculating the weight of each channel is as follows:

[0141] The first Features of each channel After being mapped to scalar scores using an MLP (Support Vector Machine), the scores are... :

[0142]

[0143] Then Normalized to weights using the Softmax weighting mechanism :

[0144]

[0145] in, and .

[0146] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0147] This indicates the influence of features in the state log on power consumption fluctuations. After calculating the prediction error during training, the backpropagation algorithm is used to update the MLP parameters to update... .

[0148] Specific implementation method eight: Combination Figure 3 This embodiment is described below. The difference between this embodiment and any one of specific embodiments one through seven is that the specific process of step five is as follows:

[0149] (1) When the number of computing nodes in the cluster that need to predict power consumption is 1, the power prediction model deployed on the computing nodes is the standard Transformer.

[0150] (2) When the number of computing nodes in the cluster that need to predict power consumption is greater than 1, the power prediction model deployed on each computing node is an improved Transformer.

[0151] The improved Transformer specifically refers to a Transformer model that replaces the attention mechanism in the standard Transformer encoder with a graph attention mechanism.

[0152] The input to the improved Transformer is: the fusion characteristics of the compute node where the improved Transformer resides and the adjacency matrix between all compute nodes in the cluster;

[0153] The adjacency matrix between all computing nodes in the cluster is calculated in method 1, method 2, or method 3.

[0154] Method 1: If the i-th node communicates with the l-th node, then the element in the i-th row and l-th column of the adjacency matrix... =1, otherwise the element in the i-th row and l-th column of the adjacency matrix =0, meaning the adjacency matrix is ​​a binary adjacency matrix composed of 0s and 1s;

[0155] Method 2: The element in the i-th row and l-th column of the adjacency matrix , = , This represents the inverse bandwidth ratio between the i-th node and the l-th node (obtained based on inter-node network latency tests (e.g., Ping or Iperf)).

[0156] Method 3: To adapt to heterogeneous distributed clusters, the element in the i-th row and l-th column of the adjacency matrix... for:

[0157]

[0158] in, This represents the communication frequency between the i-th node and the l-th node;

[0159] This indicates the total number of nodes in the cluster.

[0160] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0161] By integrating the standard Transformer encoder with a graph attention mechanism, temporal and structural feature representations can be extracted. The constructed adjacency matrix 𝐴 is used as a mask for the attention module in the improved Transformer encoder. When executing the attention mechanism, the model only calculates attention weights for neighboring nodes that are structurally connected to the current node, masking unconnected nodes and limiting the scope of attention diffusion. When calculating cross-node attention weights, paths are pruned according to the adjacency matrix, ultimately outputting the embedded features of each node under both topological and temporal semantics. When only one computing node needs to predict power consumption, the power prediction model only needs to be deployed on that computing node. When multiple computing nodes need to predict power consumption, the power prediction model is deployed on each of the computing nodes that need power consumption prediction. The input of the current power prediction model is the features of the computing node where the current model resides and the adjacency matrix.

[0162] Specific Implementation Method Nine: Combining Figure 4 This embodiment is described below. This embodiment differs from specific embodiments one through eight in that the method further includes step six, the specific process of which is as follows:

[0163] Based on the output of the last attention layer of the power consumption prediction model, the importance ranking of each dynamic state parameter is obtained, and optimization suggestions are generated based on the importance ranking results.

[0164] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0165] This invention introduces an attention interpretation mechanism and a feature importance modeling mechanism. The encoder uses a multi-head attention mechanism to process the input feature sequence, and the model can output the weighting coefficients of each feature channel (such as CPU utilization, memory usage, IO rate, power value, etc.) in the last attention module. These attention weights are aggregated and processed, combined with internal model gradient information or SHAP-like feature sensitivity metrics, to generate a set of feature contribution rankings. Based on the top-ranked "key power consumption factors," the system automatically matches a tuning mapping library to generate natural language suggestions. This invention uses theoretical analysis or engineering experience to generate a tuning mapping library. For example, if the model identifies excessively high CPU utilization as the dominant power consumption factor, the mapped tuning strategy would suggest reducing the number of parallel tasks or threads.

[0166] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that the loss function used during the training of the power consumption prediction model is... for:

[0167]

[0168] in, The label-smoothed cross-entropy loss function represents the multi-class cross-entropy loss function.

[0169] This refers to Huber Loss;

[0170] and It's a hyperparameter.

[0171] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0172] and These are hyperparameters representing the "task uncertainty" of each task. The model automatically learns these two parameters during training, and the weights are "adaptively and dynamically adjusted." Compared to the traditional joint loss method with statically set weights, the loss function designed in this invention has stronger adaptive capabilities and is suitable for big data operation scenarios with drastic power consumption fluctuations and diverse optimization suggestion strategies.

[0173] During the model training phase: the system uses the aforementioned joint loss function to perform gradient backpropagation, iteratively optimizing the encoder and multi-task output branch parameters. Specifically, this includes:

[0174] (1) Power prediction task branch Output the active power value of the target node at the next moment;

[0175] (2) Optimize strategy branch Generate executable dynamic tuning strategies.

[0176] During the inference phase: the system only performs forward prediction and does not involve loss function or weight calculation; the output includes: power consumption prediction value. and recommended tuning suggestions text .

[0177] The trained power consumption prediction model can be deployed as an embedded module into the scheduling system of a big data platform to receive system status in real time, predict future power consumption trends, and automatically generate parameter optimization suggestions.

[0178] Furthermore, the method of this invention also has a visualization function: such as Figure 5The image shows the task submission interface, which allows users to submit tasks through a visual interface. It provides two buttons for users to choose whether to perform task power consumption prediction and node status control. Users can specify the start time and execution frequency of the task on the host, and also set a task description. Before submitting a task, operations personnel can set fault or anomaly thresholds for any compute node in the cluster. Therefore, log configuration, alarm configuration, trigger configuration, and status monitoring are configured. Log configuration is used to obtain log and alarm records from the system; alarm configuration is used to select alarms based on trigger rules; trigger configuration is used to define conditions (such as resource exceeding limits) to trigger alarms; and status monitoring allows users to choose whether to enable node status monitoring during operation. Figure 6 The interface shown is for node status optimization, providing targeted energy efficiency optimization suggestions for both master and worker nodes in the cluster. By adjusting key parameters such as CPU, memory, storage, and network, it balances performance and power consumption (for both master and worker nodes), allowing operations personnel to more effectively manage task configurations. A node power consumption monitoring interface is also included, specifically:

[0179] (1) Task execution log: Records the execution process of each task in the cluster, including host, task type, input data size, duration, start date, throughput and real time.

[0180] (2) Cluster node power meter data: records the real-time power consumption of the node collected from the power meter.

[0181] (3) Average CPU utilization of the cluster: reflects the overall saturation of computing resources in the cluster.

[0182] (4) Application information: Displays the name, status, and time of the running task, as well as the current cluster status.

[0183] (5) Monitoring of core indicators: The four indicators that have a significant impact on power consumption are monitored: CPU frequency, CPU utilization, IO wait time, and total I / OPS.

[0184] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for monitoring and predicting power consumption of computing nodes, characterized in that, The method specifically includes the following steps: Step 1: Obtain the static configuration parameters of compute node i, use a power meter to collect the power consumption of compute node i during the execution of the workload in real time, and collect the dynamic status parameters of compute node i during the execution of the workload in real time through Linux commands. Step 2: Align the timestamps of the dynamic state parameters of compute node i with the timestamps of the power consumption data of compute node i to obtain the synchronized time series data of the dynamic state parameters and power consumption of compute node i. The time-series data of the power consumption of computing node i is as follows: , For real numbers, For the first Active power at each sampling point The sampling length; The time series data of the dynamic state parameters of node i are: , , For the first The dynamic state parameter sequence of each sampling point For the first CPU utilization per sampling point For the first Memory usage per sampling point For the first Network input rate at each sampling point For the first Network output rate per sampling point; Step 3: Process time series data Normalization and differencing are performed to obtain the power consumption differential sequence data; the time series data... The data for each dimension are normalized and differentially processed to obtain differential sequence data for CPU utilization, memory usage, network input rate, and network output rate. Align the differential sequence data of power consumption, CPU utilization, memory usage, network input rate, and network output rate of computing node i to form the input tensor matrix of computing node i. Step 4: Construct a static attribute vector based on the static configuration parameters of compute node i. and static attribute vector Mapped into a dense vector; The fusion features of computing node i are obtained based on the input tensor matrix and dense vector of computing node i. Step 5: Use the fusion features of computing node i as input to the power prediction model deployed on computing node i, and output the power prediction result of computing node i through the power prediction model.

2. The method for monitoring and predicting power consumption of a computing node according to claim 1, characterized in that, The static configuration parameters specifically include the number of CPU cores, memory size, hard disk type, and node role.

3. The method for monitoring and predicting power consumption of a computing node according to claim 2, characterized in that, The dynamic status parameters specifically include CPU utilization, memory usage, network input rate, and network output rate.

4. The method for monitoring and predicting power consumption of a computing node according to claim 3, characterized in that, The time series data Normalization and differencing are performed to obtain the power consumption differential sequence data; specifically: in, Power consumption time series data The mean of all data in the dataset; Power consumption time series data The standard deviation of all data in the dataset; Power consumption time series data In The corresponding normalized value; Then the first in the differential sequence data of power consumption Value for: 。 5. The method for monitoring and predicting power consumption of a computing node according to claim 4, characterized in that, The static attribute vector is constructed based on the static configuration parameters of computing node i. Specifically: Where c is the number of CPU cores in computing node i; m is the memory size of computing node i; d represents the hard disk type of compute node i; role is the role of compute node i; Encode for One-Hot; It is a static attribute vector; By embedding Mapped to a dense vector: in, Represents a dense vector; This indicates mapping processing.

6. The method for monitoring and predicting power consumption of a computing node according to claim 5, characterized in that, The process of obtaining the fusion features of computing node i based on its input tensor matrix and dense vector is as follows: Step 41: in, This represents the first element in the input tensor matrix of computation node i. Features corresponding to each sampling point; Indicates an embedded encoder; Indicates the first The location code of each sampling point; Indicates intermediate variables; Step 42: According to Obtain the computation node i at the th Fusion features of each sampling point: in, express The Middle Characteristics of each channel; Indicates the total number of channels; Indicates that the computation node i is at the 1st epoch. Fusion features of each sampling point; Indicates the first The sampling point of the first sampling point The weight of each channel; Step 43: Using the methods in Step 41 and Step 42, process the features corresponding to each sampling point in the input tensor matrix of computing node i to obtain the fusion features of computing node i at each sampling point, and use the fusion features of each sampling point to form the fusion features of computing node i.

7. The method for monitoring and predicting power consumption of a computing node according to claim 6, characterized in that, The first The sampling point of the first sampling point The method for calculating the weight of each channel is as follows: The first Features of each channel After MLP mapping, it is converted into a scalar fraction. : Then Normalized to weights using the Softmax weighting mechanism : in, and .

8. The method for monitoring and predicting power consumption of a computing node according to claim 7, characterized in that, The specific process of step five is as follows: (1) When the number of computing nodes in the cluster that need to predict power consumption is 1, the power prediction model deployed on the computing nodes is the standard Transformer. (2) When the number of computing nodes in the cluster that need to predict power consumption is greater than 1, the power prediction model deployed on each computing node is an improved Transformer. The improved Transformer specifically refers to a Transformer model that replaces the attention mechanism in the standard Transformer encoder with a graph attention mechanism. The input to the improved Transformer is: the fusion characteristics of the compute node where the improved Transformer resides and the adjacency matrix between all compute nodes in the cluster; The adjacency matrix between all computing nodes in the cluster is calculated in method 1, method 2, or method 3. Method 1: If the i-th node communicates with the l-th node, then the element in the i-th row and l-th column of the adjacency matrix... =1, otherwise the element in the i-th row and l-th column of the adjacency matrix =0, meaning the adjacency matrix is ​​a binary adjacency matrix composed of 0s and 1s; Method 2: The element in the i-th row and l-th column of the adjacency matrix , = , This represents the inverse bandwidth ratio between the i-th node and the l-th node; Method 3: The element in the i-th row and l-th column of the adjacency matrix for: in, This represents the communication frequency between the i-th node and the l-th node; This indicates the total number of nodes in the cluster.

9. The method for monitoring and predicting power consumption of a computing node according to claim 8, characterized in that, The method further includes a sixth step, the specific process of which is as follows: Based on the output of the last attention layer of the power consumption prediction model, the importance ranking of each dynamic state parameter is obtained, and optimization suggestions are generated based on the importance ranking results.

10. The method for monitoring and predicting power consumption of a computing node according to claim 9, characterized in that, The loss function used during the training of the power consumption prediction model for: in, Represents the cross-entropy loss function; This refers to Huber Loss; and It's a hyperparameter.

Images