Estimation device and estimation program

Abstract

An estimation device 10 comprises a workload power consumption estimation unit 103 that links a server power consumption estimation model for learning at least the relationship between a resource use state of a server and power consumption of the server to estimate power consumption on the server, and a workload additional learning model for learning information on power consumption of the workload on the server, and that calculates an estimated value of power consumption when a prescribed workload operates on a server of a movement destination.

Description

Estimation device and estimation program

[0001] The present disclosure relates to an estimation device and an estimation program.

[0002] By migrating workloads such as virtual machines (VMs) running on servers in a data center to another data center, we hope to improve the renewable energy utilization rate at each data center.

[0003] To achieve this, it is necessary to estimate the power consumption for each workload. However, unlike the power consumption of the server itself, the power consumption of a workload cannot be directly measured using a power distribution unit (PDU) or the like.

[0004] Therefore, in Non-Patent Document 1, it is assumed that the power consumption of a server is proportional to the CPU usage rate, and a relational expression between each CPU usage rate and the power consumption per 1% of CPU usage rate for the host OS and each VM is prepared for each server, and the power consumption of each VM on the server before migration is estimated by calculating the power consumption from these relational expressions. Then, based on the estimation result, the power consumption of each VM on the destination server is estimated from the power saving ratio of the server before and after the migration.

[0005] Yoshikazu Nakamura and six others, "VM Power Consumption Prediction Method Considering VM Migration," 2022 General Conference, B-14-6, March 15, 2022

[0006] However, even when a workload is moved between servers of the same model and specifications (when the power saving ratio is 1:1), the power decrease at the source server and the power increase at the destination server due to the workload movement may not be the same value. Therefore, the method using the power saving ratio between servers as shown in Non-Patent Document 1 has the problem of low estimation accuracy.

[0007] The present disclosure has been made in consideration of the above circumstances, and an object of the present disclosure is to provide a technology that can estimate the power consumption of a workload with high accuracy.

[0008] An estimation device according to one aspect of the present disclosure includes an estimation unit that calculates an estimate of the power consumption when a specified workload runs on a destination server by linking a server power consumption estimation model that learns at least the relationship between the server's resource usage status and the server's power consumption to estimate the power consumption of the server, and an additional learning model for workloads that learns information about the power consumption of the workloads on the server.

[0009] An estimation program according to an aspect of the present disclosure causes a computer to function as the estimation device.

[0010] According to the present disclosure, it is possible to provide a technique that can estimate the power consumption of a workload with high accuracy.

[0011] FIG. 1 is a diagram showing an overview of ideas 1 to 3. FIG. 2 is a diagram showing the load state of a workload, etc. FIG. 3 is a diagram showing the characteristics of a workload. FIG. 4 is a diagram showing the configuration of a system according to this embodiment. FIG. 5 is a diagram showing the processing flow of a power consumption estimation method. FIG. 6 is a diagram showing the hardware configuration of an estimation device.

[0012] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the description of the drawings, the same parts are designated by the same reference numerals and the description thereof will be omitted.

[0013] In this embodiment, the power consumption of a workload is estimated by machine learning. To achieve high-accuracy estimation in a short time, the following measures 1 to 3 are implemented. FIG. 1 is a diagram showing an overview of measures 1 to 3.

[0014] Idea 1: In this embodiment, the workload power consumption estimation model is considered to be a combination of a server power consumption estimation model and a workload additional learning model.

[0015] The data required to generate a server power consumption estimation model can be obtained in advance in large quantities as actual measurement data from servers, etc. Therefore, in learning phase 1, the relationship between the server resource usage status, etc. and the server power consumption is learned, and a server power consumption estimation model is generated to estimate the power consumption of the server (including the power consumption of the workload on the server) using various information output from the server.

[0016] The measured data includes, for example, resource usage (CPU usage rate, memory usage rate, etc.), workload resource usage rate, power consumption of the server, server specifications, server temperature, and temperature around the server.

[0017] In learning phase 2, a workload additional learning model is generated to learn information about the server, particularly information about the power consumption of the workload on the server. Specifically, the workload additional learning model inputs the estimated value of the power consumption of the workload on the server estimated by the server power consumption estimation model, the power consumption value of the workload calculated from the actual measured value of the power consumption of the server (hereinafter simply referred to as the actual measured value of the power consumption of the workload), and workload characteristic information (device 3 described below) to learn the relationship between them.

[0018] For example, the workload incremental learning model learns the relationship between the estimated power consumption of a workload and the actual measured power consumption of the workload, the relationship between the estimated power consumption of a workload and the characteristic value of the workload, the relationship between the estimated power consumption of a workload and the actual measured power consumption of the workload and the characteristic value of the workload, the relationship between the estimated power consumption of the same workload on different servers, and the relationship between the estimated power consumption of different workloads on the same server.For example, it learns the error between the estimated value and the actual measured value due to subtle differences between servers related to manufacturing lots, etc.

[0019] In the inference phase, the server power consumption estimation model and the workload incremental learning model are linked to estimate the power consumption of the workload on the server. Specifically, the power consumption of a given workload when it runs on the destination server is estimated.

[0020] As a method for linking models, for example, techniques such as "transfer learning" and "fine tuning" used in the field of natural language processing are used. In other words, in the server power consumption estimation model, the workload-related learning, which is part of the server-related learning, is entrusted to the workload-related additional learning model.

[0021] When there is a large amount of training data, such as resource usage data, that can be acquired from a server, and a small amount of training data that cannot be acquired on a workload-by-workload basis, by using the method described above, it is possible to narrow down the range of model parameter adjustments using training data to the workload-specific additional training model. As a result, it is possible to estimate the power consumption of a workload with high accuracy in a short time.

[0022] Technique 2: As explained in Technique 1, the workload additional learning model inputs the actual measured values ​​of the workload's power consumption as training data.

[0023] However, as shown in Fig. 2(a), the load state of a workload generally changes sequentially. Also, as shown in Fig. 2(b), it takes a certain amount of time for the power consumption of servers A and B to stabilize before and after the workload transfer. Therefore, it is difficult to determine at what point the actual measured power consumption of the server before and after the workload transfer becomes stable.

[0024] Therefore, in order to obtain effective training data (measured values ​​of appropriate power consumption), the server power consumption estimation model estimates the power consumption of the workload on the server and also estimates the time it takes for the server's power to stabilize (see Figure 1). This time can be estimated, for example, based on measured data on the server's power consumption and measured data on the server's resource usage.

[0025] The workload-specific learning model uses the actual measured values ​​of power consumption after the estimated power stabilization time has elapsed as training data, which increases the value of the actual measured values ​​of power consumption as training data and enables the workload power consumption to be estimated more accurately in a shorter time.

[0026] Tip 3: Each workload has its own unique characteristics, which affect power consumption. These characteristics include the number of threads, availability of GPU acceleration, etc.

[0027] For example, workloads that can run on multiple threads will have different power consumption because the number of parallel processes will differ depending on the server's operating environment. For example, workload WL1 shown in Figure 3(a) cannot increase the number of threads due to the application's design, while workload WL2 shown in Figure 3(b) can handle four threads. Therefore, even if the destination server is the same, if that server can simultaneously execute four or more threads, workload WL2 will consume more power.

[0028] In addition, workloads that can be offloaded to the GPU can be processed in parallel in environments where the GPU is available in addition to the CPU, which results in differences in power consumption. Depending on the application's design, the GPU may not be available or only certain GPUs may be available, which results in differences in power consumption.

[0029] Therefore, the workload-specific learning model uses workload characteristic information as training data (see Figure 1). This allows us to absorb differences in power consumption due to workload-specific characteristics and estimate the power consumption of a workload with higher accuracy in a short time.

[0030] FIG. 4 is a diagram showing the configuration of the system 1 according to this embodiment.

[0031] The system 1 includes an estimation device 10 that estimates the power consumption of workloads (VMs, containers, etc.) on each of two servers A and B, and an execution device 20 that determines and executes migration of workloads between servers A and B based on the estimation results. The number of servers may be two or more.

[0032] The estimation device 10 includes a resource and power consumption collection unit 101 , a server power consumption estimation model generation unit 102 , a workload power consumption estimation unit 103 , a weighting factor preprocessing unit 104 , and a weighting factor adjustment unit 105 .

[0033] The resource and power consumption collection unit 101 has a function of collecting information on the resource usage status (CPU usage rate, memory usage rate, etc.) of each of the servers A and B. For example, the resource and power consumption collection unit 101 collects from the server A the CPU usage rate of the host OS running on the server A and the CPU usage rates of the workloads running on the server A (CPU usage rate of VM1, CPU usage rate of VM2, CPU usage rate of container 1, etc.).

[0034] The resource and power consumption collection unit 101 also has a function of collecting information on the power consumption of each server from each PDU of the servers A and B. For example, the resource and power consumption collection unit 101 collects the power consumption of server A from the PDU of server A.

[0035] The resource and power consumption collection unit 101 has a function of collecting the specification information of each of the servers A and B from a specification information storage DB.

[0036] The server power consumption estimation model generation unit 102 has the function of inputting the resource usage status of each server A and B, the power consumption of each server itself, specification information of each server, the temperature of each server, the ambient temperature of each server, etc. as learning data, learning the relationship between the resource usage status of each server A and B, the power consumption of each server itself, specification information of each server, the temperature of each server, the ambient temperature of each server, etc., and generating a server power consumption estimation model for estimating the power consumption of the server.

[0037] Specifically, the server power consumption estimation model estimates the power consumption of the workloads on servers A and B based on the resource usage status of the workloads on servers A and B. For example, the server power consumption estimation model estimates the power consumption when a specific workload runs on a destination server. The server power consumption estimation model also estimates the time it takes for the power to stabilize on the destination server after the workload has been moved.

[0038] The workload power consumption estimation unit (estimation unit) 103 inputs the resource usage status of the workload on servers A and B as inference data, and by linking the server power consumption estimation model with the workload additional learning model, estimates the power consumption of the workload on servers A and B (for example, the power consumption when a specified workload operates on the destination server) from the resource usage status of the workload on servers A and B input as inference data, and has the function of estimating the time it takes for the power to stabilize on the destination server after the workload is moved.

[0039] Here, the server power consumption estimation model and the workload additional learning model will be explained again.

[0040] The server power consumption estimation model works in conjunction with the workload incremental learning model, which is used to learn information about the power consumption of workloads on a server, to estimate the power consumption of the workloads running on the server and the time it takes for the server's power to stabilize. As described above, the two models can be linked using techniques such as "transfer learning" and "fine tuning."

[0041] The workload additional learning model inputs the estimated value of the workload power consumption on the server estimated by the server power consumption estimation model, the workload power consumption value (actual measured value of the workload power consumption) calculated from the actual measured value of the server power consumption, which is the actual measured value after the time has passed until the power on the server stabilizes after the workload is moved, and the workload characteristic information stored in the workload characteristic information storage DB, and learns the relationship between them.

[0042] When the server power consumption estimation model is not linked to the workload additional learning model, it simply estimates the power consumption of the workload on the server based on the resource usage status of the workload, etc., but when linked to the workload additional learning model, it estimates the power consumption of the workload taking into account the characteristics of the workload and the actual measured value of the power consumption of the transferred workload, etc.

[0043] By linking this with an additional learning model for workloads that has been trained using training data such as parameters related to the characteristics of the workload and the actual measured values ​​of the power consumption of the transferred workload, it is possible to estimate the power consumption of the workload with high accuracy in a short time.

[0044] The weighting coefficient preprocessing unit 104 has the function of obtaining an estimated time until power becomes stable on the server to which the workload is to be moved from the workload power consumption estimating unit 103, and instructing the resource / power consumption collecting unit 101 to collect resource usage status, etc. after the power stabilization time has elapsed.

[0045] The weighting coefficient adjustment unit 105 has a function of adjusting the weighting coefficients of the workload additional learning model using the resource usage status and the like collected by the resource and power consumption collection unit 101 as training data based on instructions from the weighting coefficient preprocessing unit 104. For example, the weighting coefficient adjustment unit 105 changes the weighting coefficients of the workload additional learning model so as to reduce the error between the estimated value and the actual measured value of the power consumption of the workload.

[0046] The execution device 20 has a function of checking the upper limit of power consumption of the destination server B and determining whether the power consumption of server B, plus the estimated power consumption of the workload to be moved, exceeds the upper limit.

[0047] The execution device 20 also has a function for determining whether or not to migrate the workload to be migrated based on the determination result. If the upper limit value is not exceeded, the execution device 20 determines that the workload to be migrated is migrated and migrates it from server A to server B. If the upper limit value is exceeded, the execution device 20 determines that the workload to be migrated is migrated. Note that this determination method is an example, and other determination methods may be used.

[0048] FIG. 5 is a diagram showing a processing flow of the power consumption estimation method.

[0049] Step S1: The resource and power consumption collection unit 101 collects the resource usage status (CPU usage rate, memory usage rate, etc.) of each server A and B, collects the power consumption of each server body from each PDU of servers A and B, and collects the specification information of each server A and B from the specification information storage DB.

[0050] Step S2: The server power consumption estimation model generation unit 102 learns the relationship between the resource usage status of each server A and B, the power consumption of each server itself, the specification information of each server, the temperature of each server, the ambient temperature of each server, etc., and generates a server power consumption estimation model to estimate the power consumption of the workload on servers A and B from the resource usage status of the workload on servers A and B, and to estimate the time it will take for the power to stabilize on the server to which the workload has been moved.

[0051] Step S3: The workload power consumption estimation unit 103 collects the resource usage status of each server A and B, collects the power consumption of each server body from each PDU of servers A and B, and collects the specification information of each server A and B from the specification information storage DB.

[0052] Then, the workload power consumption estimation unit 103 links the server power consumption estimation model with an additional learning model for workloads that learns the relationship between the estimated and actual measured power consumption values ​​of the workload and the workload's characteristic information, calculates an estimated value of the power consumption when the workload to be moved runs on the destination server B from the CPU usage rate of the workload to be moved running on the source server A, and calculates an estimated time until the power on server B stabilizes after the workload to be moved is moved.

[0053] Step S4: The workload power consumption estimator 103 transmits an estimate of the power consumption of the workload to be moved when it operates on the destination server B to the execution unit 20. The execution unit 20 determines whether the workload to be moved can be moved to server B, and if it determines that the workload can be moved, moves the workload to be moved from server A to server B.

[0054] Step S5: The workload power consumption estimator 103 transmits an estimated time required for power to stabilize on server B after the migration of the workload to be moved to the weighting coefficient preprocessor 104. The weighting coefficient preprocessor 104 instructs the resource and power consumption collector 101 to collect the resource usage status of server B and the power consumption of servers A and B after the estimated time (power stabilization time) has elapsed.

[0055] Step S6: After the workload to be moved has been moved from server A to server B and the estimated time has elapsed, the resource and power consumption collection unit 101 collects the resource usage status of the workload to be moved from server B and collects the power consumption of each server body from each PDU of servers A and B.

[0056] Step S7: The weighting coefficient adjustment unit 105 adjusts the weighting coefficient of the workload additional learning model using the resource usage status of the workload to be moved on server B collected in step S6 and the actual measured values ​​of power consumption on servers A and B as training data.

[0057] For example, the weighting coefficient adjustment unit 105 calculates an optimal value of the weighting coefficient used in the machine learning of the workload additional learning model so as to reduce the difference (error) between the estimated power consumption value estimated in step S3 and the actual measured power consumption value collected in step S6, and changes the value of the weighting coefficient to the optimal value. At this time, the workload additional learning model is re-learned using the actual measured power consumption values ​​of servers A and B after the estimated time has elapsed, which were collected in step S6.

[0058] By repeating steps S1 to S7, the estimation result of the power consumption of the workload can be improved.

[0059] As described above, according to this embodiment, the estimation device 10 is equipped with a workload power consumption estimation unit 103 that links a server power consumption estimation model for learning at least the relationship between the server's resource usage status and the server's power consumption to estimate the power consumption of the server, and an additional learning model for workloads for learning information regarding the power consumption of workloads on the server, to calculate an estimated value of the power consumption when a specified workload runs on a destination server, thereby providing a technology that can estimate the power consumption of a workload with high accuracy.

[0060] The present disclosure is not limited to the above-described embodiments, and various modifications are possible within the scope of the present disclosure.

[0061] The estimation device 10 of the present embodiment described above can be realized, for example, by using a general-purpose computer system including a CPU 901, a memory 902, a storage 903, a communication device 904, an input device 905, and an output device 906, as shown in Fig. 6. The memory 902 and the storage 903 are storage devices. In the computer system, the CPU 901 executes a predetermined program loaded onto the memory 902, thereby realizing each function of the estimation device 10.

[0062] The estimation device 10 may be implemented by one computer, or by multiple computers, or may be a virtual machine implemented on a computer.

[0063] The program for the estimation device 10 can be stored in a computer-readable recording medium such as a HDD, SSD, USB memory, CD, or DVD. The computer-readable recording medium is, for example, a non-transitory recording medium. The program for the estimation device 10 can also be distributed via a communication network.

[0064] REFERENCE SIGNS LIST 1 System 10 Estimation device 20 Execution device 101 Resource and power consumption collection unit 102 Server power consumption estimation model generation unit 103 Workload power consumption estimation unit 104 Weighting coefficient preprocessing unit 105 Weighting coefficient adjustment unit 901 CPU 902 Memory 903 Storage 904 Communication device 905 Input device 906 Output device

Claims

1. An estimation device comprising: an estimation unit that links a server power consumption estimation model for estimating power consumption at a server by learning at least the relationship between the server's resource usage status and the server's power consumption, and an additional learning model for workloads for learning information about the power consumption of workloads at the server, to calculate an estimate of the power consumption when a specified workload is operated on a destination server.

2. The estimation device of claim 1, wherein the server power consumption estimation model estimates the time until power stabilizes at the destination server to which the workload has been moved, and the workload additional learning model learns using actual measured values ​​of power consumption at the destination server after the estimated power stabilization time has elapsed.

3. The estimation device of claim 1, wherein the workload additional learning model is trained using workload characteristic information, and the estimation unit links the server power consumption estimation model with the workload additional learning model to calculate an estimate of the power consumption when a specified workload is running on a destination server.

4. An estimation program that causes a computer to function as the estimation device according to claim 1.

Images