CN-115129475-B - Energy efficiency optimization method and system for server non-perception computing platform

CN115129475BCN 115129475 BCN115129475 BCN 115129475BCN-115129475-B

Abstract

The utility model provides a server does not have perception computing platform energy efficiency optimization method and system, through gathering the characteristic information of function on the server does not have perception platform at initialization stage and operation stage, combine wherein the function that the characteristic is the same and constitute a plurality of synchronization function sets after, corresponding distribution to have idle computing resource's function processing node and deployment to corresponding processing core, the characteristic information of updating is used for realizing cycle energy efficiency management again. The invention provides the optimal operation frequency of the power synchronization mechanism synchronization function by identifying the difference of the optimal operation frequency of the server non-perception calculation function and the time sequence difference of the optimal operation frequency of the function, thereby ensuring that the function can operate according to the optimal operation frequency in the whole life cycle.

Inventors

LI CHAO
ZHANG LU
XU YECHEN
WANG XINKAI
Leng Jingwen
CHEN QUAN
GUO MINYI

Assignees

上海交通大学

Dates

Publication Date: 20260512
Application Date: 20220712

Claims (7)

1. The power synchronization-based server non-perception platform energy efficiency management method is characterized by comprising the steps of collecting characteristic information of functions on a server non-perception platform in an initialization stage and an operation stage, combining functions with the same characteristics and forming a plurality of synchronization function sets, correspondingly distributing the functions to function processing nodes with idle computing resources and disposing the function processing nodes on corresponding processing cores, and updating the characteristic information for realizing cycle energy efficiency management; The characteristic information comprises the optimal operation frequency, the operation time and the input data size of the function; The characteristic information comprises the optimal operation frequency < > of the function in the initialization stage and the operation time stage And the runtime of the function in the initialization phase and the runtime phase < > And the Input data size Input of the function; The synchronization function is to acquire the operation information of the current function according to the collected history characteristic information of the function, wherein the function has an initialization stage and an optimal operation frequency of the operation stage Obtaining by querying historical data, predicting the running time of the function by a linear regression model, and combining the functions with the same optimal running frequency and running time into a synchronous function set to form a quadruple < Characterization; the deployment refers to deploying the synchronization function to an idle CPU core on a function processing node, setting the frequency of the idle CPU core as the optimal operation frequency of the initialization stage of the synchronization function, and then monitoring and managing the working state of the CPU core.
2. The power synchronization-based server non-aware platform energy efficiency management method of claim 1, wherein the linear regression is to train a specific predictive function runtime linear regression model at each frequency according to the Input data size Input of the function Wherein a and b are the intercept and slope, respectively, in the linear regression model, 80% of the random selection function history run information is used as the training set of the model, and the remaining 20% is used as the test set to train the linear regression model.
3. The method for managing energy efficiency of server non-aware platform based on power synchronization of claim 1, wherein the working states of the CPU core are a phase transition core, a violation synchronization state core and an idle CPU core, the phase transition core is that a function running on the CPU core changes phase along with the running of the function, the function is changed from an initialization phase to a runtime phase, the required optimal running frequency of the function also changes, at this time, the node management process is responsible for adjusting the frequency of the core to be the optimal running frequency of the runtime phase of the function, the violation synchronization state core is that the function on a part of the cores cannot be operated according to a power synchronization mode due to an uncertain factor in the running process of the function, the node management process periodically monitors the phase and number of the running function on the core to judge whether the function on the CPU core is operated according to the power synchronization mode, and the idle core is that the CPU core is idle after the operation of the function is ended for the next function deployment and operation.
4. The method for managing energy efficiency of server non-aware platform based on power synchronization according to claim 1, wherein the updated characteristic information is that the node counts the number of idle CPU cores and the synchronization state of the function in the node every 100ms, records the number of idle CPU cores to facilitate the distribution of the next synchronization function, and corrects the prediction algorithm for checking the function running time according to the violation of the synchronization state, i.e. retrains the linear regression model for running time prediction according to the running history information of the function.
5. A server non-perception platform energy efficiency optimization system taking power synchronization as a characteristic for realizing the server non-perception platform energy efficiency management method according to any one of claims 1-4 is characterized by comprising a function information acquisition module, a function combination module, a function distribution module and a node management module, wherein the function information processing module acquires information such as optimal operation frequency and operation time of a function according to function identification information; the function combination module combines the functions with the same optimal operation frequency and operation time according to the operation time information of the functions under the optimal operation frequency to form a synchronous function set; the node management module carries out power synchronization management on the function distributed to the node and carries out real-time monitoring on the running state of the function, thereby ensuring the high-efficiency and synchronous running of the system; The function combination module comprises a function optimal operation frequency analysis unit, a function life cycle assessment unit and a function synchronization unit, wherein the function optimal operation frequency analysis unit obtains the optimal operation frequency of the function in an initialization stage and an operation stage from historical operation information of the function, the function life cycle assessment unit carries out life cycle assessment by using a linear regression model according to the Input data size Input of the function, the operation time and other historical information of the function to obtain the initialization time and the operation time of the function, and the function synchronization unit combines the function according to the optimal operation frequency information of the function and the life cycle information of the function to obtain a synchronization function set.
6. The server imperceptible platform energy efficiency optimization system of claim 5, wherein the function information acquisition module comprises a function identification acquisition unit, an optimal operation frequency acquisition unit and a function operation time information acquisition unit, wherein the function identification acquisition unit acquires the identification of the function according to the triggering information of the function, and the optimal operation frequency acquisition unit compares the energy efficiency of the function under different CPU core frequencies according to the performance of the function under different CPU core frequencies according to the function identification, thereby acquiring the optimal operation frequency < > of the function in an initialization stage and an operation time stage The function running time information acquisition unit acquires the initialization time and the running time of the function under different CPU core frequencies according to the Input data sizes Input of the function, and records the information; The function distribution module comprises a synchronization function grouping unit and a node selection unit, wherein the synchronization function grouping unit performs grouping distribution on the synchronization function according to the size of a synchronization function set and the number information of the deployed functions on the CPU cores, and the node selection unit inquires nodes with idle CPU cores according to the running state of the nodes on the server unaware platform and distributes the grouped functions to the nodes with the idle CPU cores.
7. The server non-aware platform energy efficiency optimization system according to claim 5, wherein the node management module comprises a function triggering unit, a system resynchronization unit and an information feedback unit, wherein the function triggering unit deploys the function on an idle CPU core and adjusts the frequency of the function according to the identification of the function, the optimal operation frequency and other information, the system resynchronization unit adjusts the frequency of the CPU core according to the information of stage change in the function operation process so that the function is executed according to the optimal operation frequency, and the information feedback unit acquires the operation state of the node, including the number of the idle CPU cores, the function synchronization state and the like, and feeds the operation state back to the function distribution module.

Description

Energy efficiency optimization method and system for server non-perception computing platform Technical Field The invention relates to a technology in the field of cloud computing, in particular to a platform energy efficiency optimization method and system for server-oriented non-perception computing (SERVERLESS COMPUTING). Background Server unaware computing is changing the deployment and management model of applications in cloud computing as an emerging cloud native architecture. The server-unaware computing mode allows the user to focus only on the application logic itself, without regard to the cumbersome tasks of resource management, etc. This greatly increases the flexibility of application deployment and the productivity of the user. Because the server has less requirements for the resources of the non-perception function, the life cycle is short, and the functions are usually deployed on the same processing core in order to improve the resource utilization rate of the server non-perception computing system. The existing power consumption management method does not consider the power consumption characteristic that server unaware computing functions are deployed on the same processor core together. Meanwhile, the existing power consumption processing method has large management overhead, and is not suitable for a server with a short life cycle and no perception calculation function. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a method and a system for optimizing the energy efficiency of a server non-perception computing platform, and the energy efficiency of a server non-perception computing function is optimized through a power synchronization mechanism. The system identifies the difference of the optimal operation frequency of the server non-perception calculation function and the time sequence difference of the optimal operation frequency of the function, proposes the optimal operation frequency of the power synchronization mechanism synchronization function, and ensures that the function can operate according to the optimal operation frequency in the whole life cycle. The invention is realized by the following technical scheme: The invention relates to a power synchronization (power synchronization) -based server non-perception platform energy efficiency management method, which is characterized in that characteristic information of functions on a server non-perception platform in an initialization stage and an operation stage is collected, functions with the same characteristics are combined to form a plurality of synchronization function sets, the synchronization function sets are correspondingly distributed to function processing nodes with idle computing resources and deployed on corresponding processing cores, and the characteristic information is updated to realize cyclic energy efficiency management. The characteristic information comprises the optimal operation frequency, the operation time and the input data size of the function. The characteristic information comprises the optimal operation frequency < Freq init,Freqexec > of the function in the initialization stage and the operation time < T init,Texec > of the function in the operation time stage and the Input data size of the function. The optimal operation frequency means that the maximum energy efficiency of the system can be achieved when the function operates at the frequency. The synchronization function is to acquire the running information of the current function according to the collected historical characteristic information of the function. Wherein the optimal operating frequency < Freq init,Freqexec > of the initialization phase and the operating phase of the function is obtained by querying the historical data. The running time of the function is predicted by a linear regression model. And combines the functions with the same optimal operating frequency and operating time into a set of synchronized functions characterized by the quadruple < Freq init,Freqexec,Tinit,Texec >. The linear regression is that a specific predictive function running time (ET) linear regression model et=a+b×input is trained at each frequency according to the Input data size Input of the function. Where a and b are the intercept and slope, respectively, in the linear regression model. 80% of the random selection function historical operating information is used as a training set of the model, and the rest 20% is used as a test set to train the linear regression model. The corresponding distribution refers to distributing the synchronization function to idle CPU cores of the processing nodes according to the quantity of the synchronization function sets and the condition of idle resources of the nodes. The deployment refers to deploying the synchronization function to an idle CPU core on a function processing node, setting the frequency of the idle CPU core as the optimal operation frequency of the initial