Power consumption estimation system, power consumption estimation method, and power consumption estimation program

Abstract

To appropriately estimate power consumption of workload.SOLUTION: In a workload power estimation server 300 for estimating power consumption of workload executed on a physical server, a power consumption estimation system 300 includes a processor 301, and the processor 301 is configured to train a plurality of short-range power models that receive a metric of the physical server as an input and output a power consumption value of the physical server in a plurality of short ranges obtained by dividing an entire power range of the physical server into a predetermined division number, train a classifier that receives the metric of the physical server as an input and outputs identification information for identifying a corresponding short range, identify the identification information for identifying the short range to be applied on the basis of a metric of the workload and the classifier, and estimate the power consumption of the workload on the basis of the metric of the workload and a short-range power model corresponding to the short range indicated by the identification information.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a technique for estimating the power consumption of a workload executed on a physical server. [Background technology]

[0002] In a typical workload execution environment, various workloads are executed side by side on a single physical server to reduce infrastructure costs, where a workload refers to software that runs on a virtual machine built on the physical server.

[0003] These workloads have diverse power consumption characteristics, with some workloads being low power consumption workloads and others being power intensive.

[0004] Being able to estimate the power consumption of individual workloads can help implement demand response and green computing measures, such as delaying power-intensive workloads to periods with energy surpluses or renewable energy where possible, or moving power-intensive workloads to data centers with energy surpluses or renewable energy supplies.

[0005] Additionally, emissions of individual workloads can be calculated from their power consumption estimates, which can help make decisions at the individual workload level to meet high-level emissions reduction targets.

[0006] Patent Literature 1 discloses a method for attributing a portion of a server's power consumption to a workload running on the server based on the load level imposed on the server by the workload. The load level can be determined based on the usage of resources such as processor, memory, disk, and network. A regression model can be used to map the load level of the workload in the form of server resource usage to a portion of the server's total power consumption. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] U.S. Patent No. 8,751,844 Summary of the Invention [Problem to be solved by the invention]

[0008] For example, the power consumption of a workload may be estimated using a single linear regression model, which is trained with a dataset that has the full range of power consumption in, for example, a server.

[0009] For example, the intercept of a single linear regression model attempts to model the mean value of the training data set and can be large if the training data is high-power data, potentially overestimating the workload's power consumption. The intercept can be interpreted as the overhead component of the workload's power consumption due to the operating system (OS) and middleware supporting the workload, increased heat dissipation due to the workload, and other factors.

[0010] In a typical data center, there are many low-power workloads and some other power-intensive workloads. However, a model trained on a dataset with power-intensive workloads will overestimate the power consumption of the low-power workloads.

[0011] The technique disclosed in US Pat. No. 6,213,669 does not address the problem of overestimation observed when the distribution of training data contains a long tail of high-power data points.

[0012] Furthermore, workload scheduling and migration decisions based on such overestimated workload power consumption estimates may not result in increased utilization of server resources and renewable energy supplies.

[0013] The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a technique that can appropriately estimate the power consumption of a workload. [Means for solving the problem]

[0014] In order to achieve the above-mentioned object, one aspect of a power consumption estimation system is a power consumption estimation system that estimates the power consumption of a workload executed on a physical server, the power consumption estimation system having a processor, which trains a plurality of short range power models, which are power models that take metrics of the physical server as input and output power consumption values ​​of the physical server, for a plurality of short ranges obtained by dividing the entire power range of the physical server into a predetermined number of divisions, trains a classifier that takes metrics of the physical server as input and outputs specific information that identifies the corresponding short range, identifies specific information that identifies the short range to be applied based on the metrics of the workload and the classifier, and estimates the power consumption of the workload based on the metrics of the workload and the short range power model corresponding to the short range indicated by the specific information. [Effects of the Invention]

[0015] According to the present invention, the power consumption of a workload can be appropriately estimated. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a diagram showing the overall configuration of a computing system according to an embodiment. [Figure 2] FIG. 2 is a configuration diagram of a workload power estimation server according to an embodiment. [Figure 3]FIG. 3 is a configuration diagram of a metric time series database according to an embodiment. [Figure 4a] FIG. 4a is a diagram illustrating the distribution of training data for server power consumption. [Figure 4b] Figure 4b illustrates the overestimation of a single model trained at long range. [Figure 5a] FIG. 5a is a diagram illustrating the division of training data for a short-range power model according to one embodiment. [Figure 5b] FIG. 5b is a block diagram of a short range power model database according to one embodiment. [Figure 6] FIG. 6 is a flowchart of a process for developing an appropriate short-range power model according to one embodiment. [Figure 7] FIG. 7 is a flowchart of a workload power consumption estimation process according to an embodiment. [Figure 8] FIG. 8 is a flowchart of a model management process according to an embodiment. [Figure 9] FIG. 9 is a diagram showing a workload energy consumption and CO2 emission visualization screen according to one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] The following description of the embodiments will be given with reference to the drawings. Note that the embodiments described below do not limit the scope of the invention as claimed, and not all of the elements and combinations thereof described in the embodiments are necessarily essential to the solution of the invention.

[0018] FIG. 1 is a diagram showing the overall configuration of an information system according to an embodiment.

[0019] The information system 10 includes a distributed computing system configured of N clusters 100 (N is 1 or more), a workload power estimation server 300 as an example of a power consumption estimation system, and a client 400. The clusters 100, the workload power estimation server 300, and the client 400 are connected via a main network 200.

[0020] The cluster 100 includes a server 101 that acts as a master (also referred to as a master server), a server 103 that acts as a worker (also referred to as a worker server), and a cluster network 102. The master server 101 and the worker server 103 may be, for example, architecturally identical physical servers, i.e., servers with identical processors, memory, storage, networks, etc. The master server 101 and the worker server 103 communicate via the cluster network 102 using a predetermined protocol.

[0021] The master server 101 manages and monitors the execution of workloads on the worker servers 103 in the cluster 100 to which it belongs. The master server 101 obtains time-stamped metrics of server power consumption and server and workload resource usage (e.g., processor usage, memory usage, storage usage, network usage, etc.) from each server via the cluster network 102, for example, at regular intervals, using an appropriate interface such as an SNMP (Simple Network Management Protocol) interface or an API (Application Programming Interface), and pushes the metrics to a metric time series database 800 (see FIG. 2 ) in the workload power estimation server 300 via the main network 200.

[0022] The worker server 103 configures one or more virtual machines and executes one or more workloads on the virtual machines.

[0023] The workload power estimation server 300 uses time-stamped metrics to train a power estimation model to estimate power consumption, or uses a pre-trained model to estimate workload power consumption in real time using time-stamped metrics.

[0024] The client 400 executes a workload in one of the clusters 100 and monitors the power consumption of the workload estimated by the workload power estimation server 300 .

[0025] Next, the workload power estimation server 300 will be described in detail.

[0026] FIG. 2 is a configuration diagram of a workload power estimation server according to an embodiment.

[0027] The workload power estimation server 300 is an example of a computer and includes a processor 301, a memory 302, a local storage 303, a network interface 304, and an input / output device 305. The processor 301, the memory 302, the local storage 303, the network interface 304, and the input / output device 305 are communicatively connected via a system bus 306.

[0028] The processor 301 is, for example, a CPU (Central Processing Unit), and fetches a program read into the memory 302 and executes various processes.

[0029] The memory 302 is, for example, a RAM (Random Access Memory), and stores the programs executed by the processor 301 and necessary data.

[0030] The network interface 304 is an interface such as a wired LAN card or a wireless LAN card, and communicates with other devices (for example, the servers 101 and 103 of the cluster 100, the client 400, etc.) via the main network 200.

[0031] The input / output device 305 provides an interface for displaying information on the console 307 and inputting information from the console 307 .

[0032] The local storage 303 is a recording device that can permanently record data, such as a hard disk drive (HDD) or a solid state drive (SSD), and stores programs and data.

[0033] The local storage 303 stores an appropriate short-range power model development program 500, a workload power estimation program 600 as an example of a power consumption estimation program, a model management program 700, a metric time series database 800, a short-range power model database 900, and a power range identification classifier database 1000.

[0034] The appropriate short-range power model development program 500 is executed by the processor 301 to develop and train an appropriate configuration for a short-range power model. The workload power estimation program 600 is executed by the processor 301 to estimate the power consumption of a monitored workload. The model management program 700 is executed by the processor 301 to manage short-range power models and classifiers. The metric time series database 800 stores time-stamped server and workload metrics. The short-range power model database 900 stores short-range power models. The power range identification classifier database 1000 stores classifiers (e.g., power range prediction models) that predict the power range of a workload for each cluster.

[0035] Next, the detailed configuration of the metric time series database 800 will be described.

[0036] FIG. 3 is a configuration diagram of a metric time series database according to an embodiment.

[0037] A metric time series database 800 is provided for each cluster 100 and stores metrics of the servers and workloads in the corresponding cluster. The metric time series database 800 stores records for each point in time. The records include fields for a timestamp 801, a CPU usage 802, a memory usage 803, a disk usage 804, a network usage 805, and a power consumption 806.

[0038] Timestamp 801 stores a timestamp indicating the time when the metric corresponding to the record was acquired. CPU usage 802 stores CPU (processor) usage at the time corresponding to the record. Memory usage 803 stores memory usage at the time corresponding to the record. Disk usage 804 stores disk (local storage) usage at the time corresponding to the record. Network usage 805 stores network usage at the time corresponding to the record. Power consumption 806 stores power consumption at the time corresponding to the record. Record 807 in the first row of the metric time series database 800 shown in Figure 3 indicates that at 12:30:50 PM, CPU usage was 20%, memory usage was 100MB, disk usage was 120kB, network usage was 400kB, and power consumption was 150W.

[0039] Next, the distribution of training data used for training a power consumption model for estimating the power consumption of a server will be described.

[0040] FIG. 4a is a diagram illustrating the distribution of training data for server power consumption.

[0041] First, the power consumption of the workload, Ppred A typical linear recovery model for estimating is expressed as shown in equation (1) below. P pred =C cpu ·U cpu +C mem ·U men +C disk ·U disk +C nw ·U nw +C ···(1) where U cpu is the CPU usage, U mem is the memory usage, and U disk is the disk usage, U nw is the network usage. cpu ,C mem ,C disk ,C nw are the model coefficients corresponding to each metric. C is the model intercept, which can be interpreted as the overhead component of the workload's power consumption, due to the operating system and middleware supporting the workload, increased thermal dissipation due to the workload, and other factors. The intercept C is empirically observed to model the average power consumption of the training data points.

[0042] As can be seen in Figure 4a, the distribution of the training data contains a long tail of high-power data points, resulting in an average power consumption 201 as shown. The large intercept value of the model results in an overestimation of the power consumption of the server and workloads, as most data points are smaller than the average power consumption 201.

[0043] Next, we describe the power consumption estimates of a single linear regression model created using training data of long range (e.g., full range) power consumption.

[0044] Figure 4b illustrates the overestimation of a single model trained at long range.

[0045] As shown in Figure 4b, the ideal power consumption estimated by the linear regression model is equal to the actual power consumption, as shown by line 202. On the other hand, if a single linear regression model is created using training data of server power consumption with the distribution shown in Figure 4a, the estimated power consumption will be as shown by line 203, due to the characteristic distribution of the training data and the tendency of the model intercept to follow the mean value of the training data.

[0046] As can be seen from lines 202 and 203, the power consumption is overestimated for most of the data points, and underestimated for the high power data points.

[0047] Next, the development of the power model according to this embodiment will be described.

[0048] FIG. 5a is a diagram illustrating the division of training data for a short-range power model according to one embodiment.

[0049] In this embodiment, to reduce overestimation, a short-range power model is developed using short-range training data. Therefore, in this embodiment, the training data is divided into several sections (power ranges). The size of each power range can be the same or different.

[0050] The data belonging to a given power range is used to build a short-range power model, which is a linear regression model that can predict power consumption within that power range. Specifically, as shown in Figure 5a, a first power range 2001 is used to build Model 1, which can estimate power for only the first power range; similarly, a second power range 2002 is used to build Model 2, which can estimate power for only the second power range; and a third power range 2003 is used to build Model 3, which can estimate power for only the third power range.

[0051] The intercept of the short-range power model used to estimate power for each power range is the average power consumption for each power range, rather than the average power consumption for the entire data distribution. This reduces overestimation within each power range, improving the accuracy of workload power estimation.

[0052] Next, the short-range power model database 900 that manages the above-mentioned multiple short-range power models will be described.

[0053] FIG. 5b is a block diagram of a short range power model database according to one embodiment.

[0054] The short range power model database 900 includes a table for each cluster. The table for each cluster includes a record of the short range power model for each power range. The record includes a power range 901 and a C cpu 902 and C mem 903 and C disk 904 and C nw The record includes fields C 905 and C 906. The power range 901 stores the power range of the short-range power model corresponding to the record. cpu 902 stores the model coefficient for the CPU usage of the short-range power model corresponding to the record. mem 903 stores the model coefficients for the memory usage of the short-range power model corresponding to the record. disk 904 stores the model coefficients for the disk usage of the short-range power model corresponding to the record. nw 905 stores the model coefficient for the network usage of the short-range power model corresponding to the record. C906 stores the intercept value of the short-range power model corresponding to the record. Record 907 in the first row of the table of cluster 1 indicates that the power range of the short-range power model corresponding to the record is 1-100W, Ccpu is 0.8, Cmen is 0.7, Cdisk is 0.65, Cnw is 0.5, and C is 52.

[0055] Next, the processing operation of the workload power estimating server 300 according to this embodiment will be described.

[0056] FIG. 6 is a flowchart of a process for developing an appropriate short-range power model according to one embodiment.

[0057] The appropriate short-range power model development process shown in Figure 6 is a process for developing an appropriate short-range power model (appropriate short-range power model) for one server in a cluster. However, in this embodiment, each server in the cluster has a uniform configuration, so the developed appropriate short-range power model can be used for any server. This appropriate short-range power model development process is executed, for example, periodically or when instructed by a user.

[0058] In step 501, the appropriate short-range power model development program 500 inputs time-stamped server resource usage data and server power consumption data to be used to train the model. The server resource usage data includes time-stamped server-level metrics for processor usage, memory usage, local storage usage, and network usage, as shown in FIG. 3. The server power consumption data includes a time-stamped total server power consumption reading minus a reading of the server's standby power consumption (server standby power). The server standby power consumption reading can be determined by a reading when the server is running without any workload.

[0059] In step 502, the appropriate short-range power model development program 500 performs preprocessing on the time-stamped data, including removing outliers or replacing missing values ​​with suitable techniques, such as interpolating or averaging metrics, or other methods. Preprocessing may include time synchronization of metrics obtained from different interfaces. Preprocessing may also include using rolling or moving averages to better capture patterns in the metrics and help remove noise from the data.

[0060] In step 503, the appropriate short range power model development program 500 calculates the maximum server power P max and the minimum server power P min Record the following.

[0061] In step 504, the appropriate short-range power model development program 500 begins processing in loop 1 (504-515) by initializing N=1 to find an appropriate number of short-range power models or an appropriate number N of power ranges (i.e., an appropriate number of sections N of the training data, an appropriate number of divisions).

[0062] In step 505, the appropriate short range power model development program 500 partitions the training data into N power ranges of equal size.

[0063] In step 506, the appropriate short-range power model development program 500 begins processing loop 2 (506-514) to find an appropriate size for each of the N power ranges of the training data by initializing the number of iterations (#iterations)=0.

[0064] In step 507, the appropriate short-range power model development program 500 trains N short-range power models using the N power ranges set in loop 1 and the size of each power range set in loop 2. As described with reference to FIG. 5a, the appropriate short-range power model development program 500 trains each short-range power model using only training data obtained from its respective power range. Here, inputs to the short-range power models are metrics related to the processor, memory, storage, and network, such as processor usage, memory usage, storage usage, and network usage, and the output from the short-range power model is the server power consumption (power consumption value). This server power consumption does not include power consumption in the standby state of the server. The short-range power model uses, for example, linear regression.

[0065] In step 508, the appropriate short-range power model development program 500 uses N power ranges and the size of each power range to train a classifier capable of predicting a power range for a given timestamp or time period. Here, the input to the classifier is processor, memory, storage, and network metrics, such as processor usage, memory usage, storage usage, and network usage, and the output from the classifier is a label (identification information) that can identify the power range. For example, if the total power range of 1-800 W is divided into three short ranges, namely, 1-200 W, 201-500 W, and 501-800 W, the label for the power range for 1-200 W is 0, the label for the power range for 201-500 W is 1, and the label for the power range for 501-800 W is 2. In this way, the classifier may be a power range prediction model used to select an appropriate short-range power model for workload power estimation. The classifier can be developed using an artificial neural network, a decision tree, or any other appropriate classification model.

[0066] In step 509, the appropriate short-range power model development program 500 tests the short-range power model developed in step 507 and the classifier developed in step 508 with test data to calculate the accuracy of server power consumption. First, the appropriate short-range power model development program 500 predicts a power range for the test data input using the classifier. Next, the appropriate short-range power model development program 500 uses the short-range power model corresponding to the predicted power range to estimate the overall power consumption using the test data input. Next, the appropriate short-range power model development program 500 uses an appropriate accuracy measurement method to compare the estimated output with the actual test data output and calculate the test accuracy rate. Here, the accuracy measurement method may be, for example, RMSE (Root Mean Square Error) or MAPE (Mean Absolute Percentage Error).

[0067] In step 510, the appropriate short-range power model development program 500 compares the test accuracy rate with a preset threshold accuracy rate. If the test accuracy rate exceeds the preset threshold accuracy rate (510: Yes), in step 512, the appropriate short-range power model development program 500 stores the short-range power model trained using data divided into N power ranges of an appropriate size (appropriate division number) in the short-range power model database 900, and stores the classifier in the power range classifier database 1000.

[0068] On the other hand, if the test accuracy rate does not exceed the threshold accuracy rate (510: No), the appropriate short-range power model development program 500 compares the number of iterations of loop 2 with a predetermined iteration threshold in step 511. The iteration threshold can be defined based on, for example, the space and time complexity of the algorithm.

[0069] If the number of iterations of Loop 2 does not exceed the predetermined iteration threshold (511: No), then in step 513, appropriate short-range power model development program 500 updates the size of the N power ranges using one of the appropriate optimization procedures. For example, if the total power range of 1-800W is divided into three short ranges, 1-200W, 201-500W, and 501-800W, the new settings may be 1-100W, 101-300W, and 301-800W.

[0070] In step 514, the appropriate short-range power model development program 500 increments (adds 1 to) the number of iterations of loop 2, and repeats loop 2 from step 507.

[0071] On the other hand, if the number of repetitions exceeds the threshold in step 511, the appropriate short-range power model development program 500 exits the processing of loop 2, increments (adds 1 to) the number N of power ranges in step 515, and repeats loop 1 from step 505.

[0072] According to this appropriate short-range power model development process, the number of appropriate power ranges and the size of the power ranges are identified, and short-range power models corresponding to the identified number of power ranges and classifiers can be registered in the database.

[0073] Next, the workload power consumption estimation process will be described.

[0074] FIG. 7 is a flowchart of a workload power consumption estimation process according to an embodiment.

[0075] The workload power consumption estimation process is executed, for example, periodically or when instructed by a user.

[0076] In step 601, the Workload Power Estimation Program 600 inputs time-stamped workload resource usage data, which includes time-stamped workload-level metrics for processor usage, memory usage, local storage usage, and network usage, as shown in FIG.

[0077] In step 602, the workload power estimator 600 performs pre-processing on the time-stamped data. For example, if the short-range power model was trained on a rolling or moving average of the training data, the pre-processing may include using the rolling or moving average over a period of time.

[0078] In step 603, since the short-range power model is trained on server-level (physical server-level) metrics, the workload power estimation program 600 applies a predetermined appropriate scaling factor to the workload resource usage metrics to scale them up to server-level metrics.

[0079] For example, the workload metric U workload The server-level metric U is calculated using the following formula (2): servermay be converted to U serve =U workload / (number of processors used by the workload)*(total number of processors in the server) (2)

[0080] In step 604, the workload power estimator 600 inputs the scaled-up workload usage metrics into a pre-trained classifier that predicts power ranges.

[0081] In step 605, the workload power estimation program 600 estimates the amount of power consumption for a given period of time using a pre-trained short-range power model that corresponds to the expected power range.

[0082] In step 606, since the short-range power model has been trained using server-level data to output server power consumption, the workload power estimation program 600 scales down the server power consumption estimated in step 605 to workload-level power consumption using an appropriate predetermined factor.

[0083] For example, the server-level power consumption P ⌒ server is calculated by using the following equation (3) to calculate the workload level power consumption prediction value P ⌒ workload can be converted to P⌒ workload =P⌒ server / (total number of processors in the server)*(number of processors used by the workload) (3)

[0084] In step 607, the workload power estimation program 600 outputs the estimated workload power consumption for the given period.

[0085] According to the above-described workload power consumption estimation process, the amount of power consumption of a workload can be appropriately estimated.

[0086] Next, the model management process will be described.

[0087] FIG. 8 is a flowchart of a model management process according to an embodiment.

[0088] A model management process is performed to ensure that the short-range power model and the classifier have a predefined accuracy rate.

[0089] In step 701, the model management program 700 performs an accuracy check at scheduled times or periodically. Here, the accuracy check uses the short-range power model and the classifier to calculate the total estimated power for a given period of time for all workloads collocated on the server being checked, and compares this with the power consumption of the server for this period (here, for example, the total power consumption of the server minus the power consumption of the standby server) to calculate the accuracy rate.

[0090] In step 702, the model management program 700 determines whether the accuracy rate exceeds a threshold accuracy rate.

[0091] As a result, if the accuracy rate exceeds the threshold accuracy rate (702: Yes), the model management program 700 in step 706 ensures that the latest versions of the short-range power model and classifier for the corresponding cluster continue to be used to estimate the power consumption of workloads running on that cluster.

[0092] On the other hand, if the accuracy rate does not exceed the threshold accuracy rate (702: No), the model management program 700 creates a new training dataset and updates the version number of the training dataset in step 703. Here, the new training dataset may be, for example, data collected in the metric time series database 800 after the last update of the short-range power model.

[0093] In step 704, the model management program 700 causes the appropriate short-range power model development program 500 to be executed using the latest version of the training dataset, thereby executing the appropriate short-range power model development process to develop a short-range power model and a classifier that are compatible with the latest version of the training dataset.

[0094] In step 705, the model management program 700 updates the version numbers of the short-range power models and classifiers developed in step 704 with the latest version of the training dataset.

[0095] In step 706, model manager 700 causes the latest version of the short-range power model and classifier for the corresponding cluster to be used to estimate the power consumption of workloads running on that cluster.

[0096] This model management process allows the short-range power model and classifier to be updated appropriately.

[0097] Next, the workload energy consumption and CO2 emission visualization screen (visualization screen) displayed by the workload power estimation server 300 will be described.

[0098] FIG. 9 is a diagram showing a workload energy consumption and CO2 emission visualization screen according to one embodiment.

[0099] The visualization screen 3000 is a screen that shows the energy (power) consumption of a specified workload over a specified period and the composition of renewable energy (RE) and Brownian energy (BE) in the energy consumption, and is displayed, for example, on the console 307 or the client 400 by the workload power estimation program 600.

[0100] The visualization screen 3000 includes a workload ID input area 3001 , a start period input area 3002 , an end period input area 3003 , an add workload button 3004 , a clear button 3005 , and a graph display area 3006 .

[0101] The workload ID input field 3001 is an area for inputting the ID of the workload (workload ID) to be displayed in the graph display area 3006. The start period input field 3002 is an area for inputting the start point of the period for which power consumption is to be measured. The end period input field 3003 is an area for inputting the end point of the period for which power consumption is to be measured. The add workload button 3004 is a button for accepting an instruction to add to the graph display area 3006 an object indicating the power consumption for the period entered in the start period input field 3002 and the end period input field 3003 for the workload having the workload ID entered in the workload ID input field 3001. The clear button 3005 is a button for accepting an instruction to delete from the graph display area 3006 the object of the workload having the entered workload ID. The graph display area 3006 is an area for displaying an object indicating the power consumption and the ratio of RE to BE for the workload whose addition has been accepted. Note that information on RE and BE corresponding to a workload can be identified by externally acquiring information on the power supplied to the server that executes the workload.

[0102] The present invention is not limited to the above-described embodiment, and can be modified appropriately without departing from the spirit of the present invention.

[0103] For example, in the above embodiments, some or all of the processing performed by the processor may be performed by a dedicated hardware circuit. Furthermore, the programs in the above embodiments may be installed from a program source. The program source may be a program distribution server or a recording medium (e.g., a portable recording medium). [Explanation of symbols]

[0104] 10...Information system, 100...Cluster, 102...Intra-cluster network, 101, 103...Server, 200...Main network, 300...Workload power estimation server, 301...Processor, 302...Memory, 303...Local storage, 304...Network interface, 305...Input / output device, 306...System bus, 307...Console, 400...Client

Claims

1. A power consumption estimation system for estimating power consumption of a workload executed on a physical server, comprising: The power consumption estimation system includes a processor, The processor: training a plurality of short range power models, which are power models that input metrics of the physical server and output power consumption values ​​of the physical server, for a plurality of short ranges obtained by dividing the entire power range of the physical server into a predetermined number of divisions; training a classifier that takes the metrics of the physical servers as input and outputs specific information that identifies the corresponding short ranges; Identifying specific information that identifies a short range to be applied based on the workload metrics and the classifier; Estimating the power consumption of the workload based on the metrics of the workload and a short-range power model corresponding to the short range indicated by the identification information. Power consumption estimation system.

2. The processor: changing the number of divisions into which the entire power range of the physical server is divided, training the short-range power model for each short range for each of the number of divisions, and calculating the accuracy rate of power consumption estimated using the short-range power model for the number of divisions, thereby detecting the number of divisions for which the accuracy rate exceeds a predetermined threshold; The detected number of divisions is determined as the number of divisions when estimating the power consumption of the workload. The power consumption estimation system according to claim 1 .

3. The processor: changing the size of the short range, training the short range power model for each short range according to the respective sizes, and calculating an accuracy rate of power consumption estimated using the short range power model, thereby detecting the size of the short range when the accuracy rate exceeds a predetermined threshold; The detected size of the short range is determined as the size of the short range when estimating the power consumption of the workload. The power consumption estimation system according to claim 1 .

4. The short-range power model outputs, as the power consumption of the physical server, the power consumption obtained by subtracting the power consumption of the physical server in a standby state from the total power consumption of the physical server. The power consumption estimation system according to claim 1 .

5. The processor: converting the workload metrics into physical server-level metrics; inputting the physical server-level metric into the classifier to identify a short range corresponding to the physical server-level metric; inputting the metrics at the physical server level into a short-range power model corresponding to the identified short range to estimate power consumption at the physical server level; Converting the estimated power consumption to a power consumption at the level of said workload. The power consumption estimation system according to claim 1 .

6. The processor: acquiring new training data including metrics and power consumption of a physical server after previously training the short-range power model; changing the number of divisions into which the entire power range of the physical server is divided, training a short-range power model for each short range for each of the number of divisions using the new training data, and calculating the accuracy rate of power consumption estimated using the short-range power model for the number of divisions, thereby detecting an appropriate number of divisions that is the number of divisions when the accuracy rate exceeds a predetermined threshold; The detected appropriate division number is determined as the division number when estimating the power consumption of the workload. The power consumption estimation system according to claim 1 .

7. The processor: Based on the power consumption of the physical server and the sum of the power consumption estimated by the short-range power model for all workloads collocated on the physical server, it is determined whether the estimation of the power consumption of the workload by the short-range power model has a predetermined accuracy, and if the estimation does not have the predetermined accuracy, new training data is acquired, and the appropriate division number is detected using the new training data, and the appropriate division number is determined as the division number when estimating the appropriate division number. The power consumption estimation system according to claim 6 .

8. The processor: Displaying the estimated power consumption of the workload The power consumption estimation system according to claim 1 .

9. A power consumption estimation method for a power consumption estimation system that estimates power consumption of a workload executed on a physical server, comprising: The power consumption estimation system includes: training a plurality of short range power models, which are power models that input metrics of the physical server and output power consumption values ​​of the physical server, for a plurality of short ranges obtained by dividing the entire power range of the physical server into a predetermined number of divisions; training a classifier that takes the physical server metrics as input to identify corresponding short ranges; Identifying a short range to apply based on the workload metrics and the classifier; Estimating the power consumption of the workload based on metrics of the workload and a short-range power model corresponding to the identified short range. Power consumption estimation method.

10. A power consumption estimation program that causes a computer to execute a process of estimating power consumption of a workload executed on a physical server, The power consumption estimation program To the computer training a plurality of short range power models, which are power models that input metrics of the physical server and output power consumption values ​​of the physical server, for a plurality of short ranges obtained by dividing the entire power range of the physical server into a predetermined number of divisions; training a classifier that takes the physical server metrics as input to identify corresponding short ranges; Identifying a short range to be applied based on the workload metric and the classifier; Estimating power consumption of the workload based on metrics of the workload and a short-range power model corresponding to the identified short range. Power consumption estimation program.

Images