CN103902016A - Server power consumption management method oriented to scene prediction - Google Patents

Abstract

The invention provides a server power consumption management method oriented to scene prediction. Firstly, historical power consumption data under a specific scene of a server system are acquired; then, the historical data are sorted and analyzed in an off-line data mining method, a prediction strategy model oriented to the application scene is established according to the data mining result, and a power consumption strategy based on the prediction strategy model is given; finally, dynamic power consumption management is performed oriented to a NodeManager. With the server power consumption management method oriented to scene prediction, effective power consumption management can be performed under various scenes where a large-scale server is deployed, not only can the multi-level energy saving results from the chip level, the infrastructure level, the software service level and the like be achieved, but also the overall cooling efficiency of a machine room can be improved, energy is greatly saved, and maintenance and operation cost is reduced.

Description

A Scenario Prediction-Oriented Server Power Consumption Management Method

technical field

The invention designs an interactive interface development technology, specifically a scene prediction-oriented server power consumption management method.

Background technique

The increase in server energy consumption is mainly reflected in the following aspects: First, the increase in CPU power consumption. For decades, the manufacturing process of CPU has been continuously improved, and the frequency has become higher and higher. While its computing power has increased rapidly, its power consumption The same growth is amazing; followed by the increase in memory and its power consumption, with the improvement of the manufacturing process, the memory capacity is getting larger and faster; the chipset and peripherals are again, due to the continuous increase in CPU and memory frequency, This requires that the chipset, bus and peripherals that work with them need to work at a higher frequency in order to give full play to their performance, and a higher frequency means more power consumption; the fourth aspect is the cooling and power supply of the computer room The energy consumption of the equipment, because the server is very sensitive to high temperature, so the computer room must be equipped with a large number of cooling equipment, and the power supply equipment in the computer room will lose part of the power during voltage conversion. Taking a petaflop supercomputer as an example, its annual electricity bill is expected to be as high as tens of millions of RMB, and the energy consumption of the petaflop supercomputer system in the future is expected to be much higher than the above estimate. Therefore, the current energy saving and consumption reduction It has gradually become a keyword of server technology.

At present, the industry's research on server energy-saving technologies mainly focuses on three aspects: chip-level energy-saving technologies, such as CPU power consumption control, CPU frequency adjustment, etc.; infrastructure-level energy-saving technologies, such as high-efficiency power supplies, intelligent temperature-controlled fans, etc.; system Level energy-saving technologies, such as cluster node hibernation based on job scheduling, energy-consuming oriented processes, and job-level migration. Although the above aspects can save part of the power consumption, they cannot improve the overall cooling efficiency of the computer room, so it cannot achieve a better effect of saving energy and reducing consumption of the server.

Therefore, how to accurately manage the actual power consumption of the server system, improve the overall energy saving and consumption reduction efficiency of the server room, and further realize an autonomous server system power consumption management mechanism with predictive processing capabilities has become an urgent need to solve The problem.

Invention content

Aiming at how to accurately manage the actual power consumption of the server system and improve the efficiency of the overall energy saving and consumption reduction of the computer room, the present invention proposes a method for managing the power consumption of the server system with a scenario-oriented prediction model based on data mining. .

The scenario prediction-oriented server power consumption management method of the present invention adopts the following technical solution to solve the technical problem: the scenario prediction-oriented server power consumption management method mines the actual power consumption of the server system in various application scenarios Specifically, a prediction model between server system power consumption and application scenarios is established, and server power consumption is managed from the perspective of prediction; the main content of the scenario prediction-oriented server power consumption management method includes: server historical data mining method , Prediction strategy model, setting and implementation of power consumption management strategy;

The scenario prediction-oriented server power consumption management method first obtains the historical power consumption data of the server system in a specific scenario, then organizes and analyzes the historical data through an offline data mining method, and establishes an application scenario-oriented prediction according to the data mining results Strategy model, and a power management strategy based on the predictive strategy model, and finally implemented for Node Manager for dynamic management of power consumption.

Further, in the server historical data mining method, the server historical data comes from historical data collected under different application scenarios of the server system, and data mining has clear scenario orientation and timeliness;

The data mining method refers to firstly preprocessing the data through data extraction, conversion, washing, integration and loading, and then analyzing and processing the data through aggregation, regression, classification algorithms and association relationships; wherein, the data mining analysis is aimed at the server The historical power consumption data collected in different application scenarios of the system is carried out offline.

Further, the establishment of the prediction strategy model is based on the results of data mining, which is a conversion model proposed for server system power consumption trends and application scenarios, and provides the conversion relationship between server system power consumption and application scenarios. Management provides predictive control strategies.

Further, the power consumption management strategy is given based on the above prediction strategy model, that is, the formulation of the power consumption management strategy is based on the prediction strategy model; the setting items of the power consumption management strategy include: strategy label, strategy type, Policy power consumption threshold, policy execution period, and policy cycle period.

Further, the implementation of the power management strategy: implement the power management strategy through the Node Manager, first set the specific management strategy to the ME of the Node Manager by the BMC through the SMBus bus management interface between the BMC and the ME, and then re- Start the BMC to enable the ME control system, CPU, and memory components to manage power consumption.

The beneficial effects of a scene prediction-oriented server power consumption management method disclosed by the present invention are:

The method of the present invention has nothing to do with specific levels of power consumption management technology, and can manage server power consumption from the perspective of prediction by establishing a predictive relationship between server system power consumption and application scenarios. Using the scene prediction-oriented server power consumption management method of the present invention, effective power consumption management can be performed in various application scenarios for large-scale server deployment. Through the scene prediction-oriented control method, not only the chip level, infrastructure Level, software business level and other multi-level energy-saving results, and can improve the overall cooling efficiency of the computer room, greatly saving energy, reducing maintenance and operation costs. In addition, the present invention also relates to offline data mining analysis, in which the accuracy of the mining algorithm is closely related to the establishment of power consumption and scene conversion models, and the power consumption management can be further improved by improving and optimizing the mining algorithm Accuracy and Validity.

This method is suitable for all IntelX86 processor platforms that support Node Manager 2.0 and above, especially for big data processing centers, high-performance cluster computing, etc. The power consumption in the application situation can better meet the requirements of enterprises to reduce energy consumption.

Description of drawings

Accompanying drawing 1 is the block diagram of the power consumption management oriented scene prediction of the present invention;

Accompanying drawing 2 is the flow chart of data mining analysis of the present invention;

Accompanying drawing 3 is the working flow chart of prediction strategy model of the present invention;

Accompanying drawing 4 is the connection diagram of parts such as BMC, ME, PSU of the present invention.

Detailed ways

The method for managing power consumption of a server oriented to scene prediction in the present invention will be further described in detail with reference to the accompanying drawings, which does not constitute a limitation to the present invention.

The scene prediction-oriented server power consumption management method of the present invention mainly involves in the design of: a scene-oriented server historical data mining method, a setting of a prediction strategy model based on power consumption and scene conversion, and a power consumption management strategy settings and implementation of power management policies. The scenario prediction-oriented server power consumption management method firstly obtains the historical power consumption data of the specific scenario of the server system, then organizes and analyzes the historical data through the data offline mining method, and then establishes the oriented The prediction strategy model of the application scenario is given, and the power management strategy based on the model is given. Finally, the management strategy is set to the Node Manager (node manager) entity ME through the BMC, and the BMC is restarted to make the ME power management effective.

In the scene-oriented server historical data mining method, the server historical data is derived from historical data collected under different application scenarios of the server system, and the data mining has clear scene orientation and timeliness;

The scene-oriented server historical data mining method refers to firstly preprocessing the data through steps such as data extraction, conversion, washing, integration and loading, and then analyzing and processing the data through aggregation, regression, classification and other algorithms and association relationships; Data mining analysis is performed offline for the historical power consumption data collected in different application scenarios of the server system.

The establishment of the forecasting strategy model is based on the results of data mining. The forecasting strategy model is a conversion model proposed for server system power consumption trends and application scenarios, and provides the conversion relationship between server system power consumption and application scenarios. Consumption management provides predictive control strategies. The application scenarios described here include: high-performance computing application scenarios, data center energy-saving scenarios, business time interval selection scenarios, energy consumption decentralized management scenarios, and the like.

In the setting of the power consumption management strategy, the power consumption management strategy is given based on the prediction strategy model, that is, the formulation of the power consumption management strategy is based on the prediction strategy model; the setting of the power consumption management strategy Items include: policy label, policy type, policy power consumption threshold, policy execution period, and policy cycle period; when setting power consumption management policies, it is subject to the following conditions: the number of policies for each independent power consumption management unit does not exceed 5 The same policy can be added, deleted, turned on, and turned off for multiple independent power management units at the same time each time. The time cycle of the power management strategy of the same independent unit cannot overlap and repeat. The power management unit type supports : System level, CPU level, memory level, PCIE level, etc.

Example:

The scene prediction-oriented server power consumption management method of the present invention supports the Intel X86 processor version of Node Manager, and the Intel X86 processor version belongs to but is not limited to RomeleySandybridgeEP processor platform, Brickland IvybridgeEX processor platform, GrantleyHaswall-EP processor platform, etc. and subsequent platforms; when implementing the forecasting strategy module, take the Intel X86 processor version of Node Manager as an example, the Node Manager is composed of hardware ME (Manager Engine) and software SPS (Service Platform Software) firmware, and Node Manager is linked with BMC through SMBus bus;

The implementation method of the power consumption management strategy described in the present invention is as follows:

To implement the power consumption management strategy through the Node Manager, first set the specific management strategy from the BMC to the ME of the Node Manager through the SMBus bus management interface between the BMC and the ME, and then restart the BMC to make the ME power consumption management strategy take effect; that is The BMC communicates with the Node Manager tool supported by the Intel X86 processor platform. According to the analysis results of the forecast strategy model, for specific application scenarios, the management strategy (power consumption type, power consumption threshold, power consumption cycle, power consumption management time) is set to ME, and the ME unit directly manages system power consumption from the system, CPU, and memory levels.

Accompanying drawing 1 is the block diagram of the power consumption management oriented scene prediction of the present invention, as shown in Fig. 1, the whole block diagram is divided into various scene data modules, data offline mining analysis module, prediction based on power consumption and scene conversion ( Strategy) model, power management strategy and modification module, strategy setting and implementation module;

The scene data module is composed of sub-modules such as performance priority scene data, energy consumption priority scene data, business concentration scene data, and energy consumption distribution scene data; data offline mining analysis is used to analyze historical data according to specific data mining algorithms and association rules and rule mining; the prediction model is established based on the results of data mining, and generates power management strategies and correction values, and finally the strategy setting and implementation module completes the execution of the entire power management.

Accompanying drawing 2 is the flow chart of data mining analysis of the present invention, as shown in Figure 2, concrete process is described as follows:

Step 1: Obtain the historical data of power consumption of the server system in each scenario;

Step 2: Data preprocessing: mainly including extraction, conversion, washing, integration, loading and other processes;

Step 3: Data mining algorithm: mainly including aggregation, regression, classification and other algorithms;

Step 4: Combining the relationship rules to carry out the relationship between the power consumption of the server system and different application occasions.

Accompanying drawing 3 is the working flowchart of forecasting strategy model of the present invention, as shown in Figure 3, concrete process is described as follows:

Step 1: Obtain the relationship between power consumption and application scenarios;

Step 2: Build a predictive model based on relationship trends;

Step 3: Determine the specific power consumption threshold of the next application scenario by the prediction model;

Step 4: Formulate a power management strategy;

Step 5: After setting ME through BMC, manage the system power consumption threshold;

Step 6: Obtain the specific data of the actual power consumption in the application scenario;

Step 7: Obtain power consumption and specific scene model correction values through data mining analysis;

Step 8: Determine whether the amendment requirements are met?

Step 9: If the correction requirements are not met (need to be corrected), go back to step 2;

Step 10: If the amendment requirements are met (no amendment required), go back to Step 5.

Accompanying drawing 4 is the connection diagram of parts such as BMC, ME, PSU (power supply) of the present invention, as shown in Figure 4, link relation is as follows:

Step 1: ME obtains power consumption management components such as CPU, memory, and system through PECI;

Step 2: BMC communicates with ME through SMBus and can send control commands;

Step 3: BMC communicates with PSU through PMBus to obtain power status and operating power;

Step 4: The server system may contain one or more nodes (2 nodes are listed in this example), and participate in the execution of the power management strategy at the same time.

In summary, the present invention improves the traditional process of server power consumption management. The traditional process mainly includes power consumption data acquisition, judging the relationship between power consumption and current system power consumption output, and adjusting the system power output; although this process improves And dynamic power management, but it is not only difficult to consider the influence of external factors of the system during the adjustment in the business process, but also inevitably cause unnecessary power loss during the power adjustment process. The scenario prediction-oriented server power consumption management method of the present invention improves the above process, adds data mining analysis and a business scenario-oriented prediction model, and sets the prediction control strategy according to the application scenario before the system runs, which can improve the system performance. efficiency of power management.

Using the scene prediction-oriented server power consumption management method of the present invention, effective power consumption management can be carried out in various application scenarios for large-scale server deployment. Level, software business level and other multi-level energy-saving results, and can improve the overall cooling efficiency of the computer room, greatly saving energy, reducing maintenance and operation costs. In addition, the present invention also relates to offline data mining analysis, in which the accuracy of the mining algorithm is closely related to the establishment of power consumption and scene conversion models, and the power consumption management can be further improved by improving and optimizing the mining algorithm Accuracy and Validity.

The above description is only an embodiment of the present invention, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims (5)

1. A scenario prediction-oriented server power consumption management method, characterized in that the server power consumption management method establishes a server system power consumption and application The prediction model between scenarios manages server power consumption from the perspective of prediction; the main content of the scenario prediction-oriented server power consumption management method includes: server history data mining method, prediction strategy model, power consumption management strategy setting and its implementation; The scenario prediction-oriented server power consumption management method first obtains the historical power consumption data of the server system in a specific scenario, then organizes and analyzes the historical data through a data mining method, and establishes a prediction strategy for application scenarios according to the data mining results model, and give a power management strategy based on the predictive strategy model, and finally implement the power management strategy for Node Manager for dynamic management of power consumption. 2. The scene prediction-oriented server power consumption management method according to claim 1, characterized in that, in the server historical data mining method, the server historical data comes from historical data collected under different application scenarios of the server system, and the data mining Have clear scene orientation and timeliness; The data mining method refers to firstly preprocessing the data through data extraction, conversion, washing, integration and loading, and then analyzing and processing the data through aggregation, regression, classification algorithms and association relationships; wherein, the data mining analysis is aimed at the server The historical power consumption data collected in different application scenarios of the system is carried out offline. 3. The scenario prediction-oriented server power consumption management method according to claim 2, characterized in that the establishment of the prediction strategy model is based on the results of data mining, and is proposed for server system power consumption trends and application scenarios The conversion model provides the conversion relationship between server system power consumption and application scenarios, and provides predictive control strategies for power consumption management. 4. The scene prediction-oriented server power consumption management method according to claim 3, wherein the power consumption management strategy is given based on the above-mentioned prediction strategy model, that is, the power consumption management strategy is formulated based on the prediction strategy model based; the setting items of the power management policy include: policy label, policy type, policy power consumption threshold, policy execution period, and policy cycle period. 5. The scene prediction-oriented server power consumption management method according to claim 4, characterized in that, in the implementation of the power consumption management strategy, the power consumption management strategy is implemented through the Node Manager, first through the communication between the BMC and the ME The SMBus bus management interface sets the specific management strategy from the BMC to the ME of the Node Manager, and then restarts the BMC to enable the ME control system, CPU, and memory components to manage power consumption.

Images