WO2020078135A1 - 资源调度方法和计算机设备 - Google Patents

资源调度方法和计算机设备 Download PDF

Info

Publication number
WO2020078135A1
WO2020078135A1 PCT/CN2019/104292 CN2019104292W WO2020078135A1 WO 2020078135 A1 WO2020078135 A1 WO 2020078135A1 CN 2019104292 W CN2019104292 W CN 2019104292W WO 2020078135 A1 WO2020078135 A1 WO 2020078135A1
Authority
WO
WIPO (PCT)
Prior art keywords
performance
temperature
power consumption
application
degradation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/104292
Other languages
English (en)
French (fr)
Inventor
陈秋林
周振坤
徐羽琼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Technologies Co Ltd
Original Assignee
Huawei Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Technologies Co Ltd filed Critical Huawei Technologies Co Ltd
Priority to EP19872996.4A priority Critical patent/EP3848776A4/en
Publication of WO2020078135A1 publication Critical patent/WO2020078135A1/zh
Priority to US17/231,268 priority patent/US20210232201A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/16—Constructional details or arrangements
    • G06F1/20—Cooling means
    • G06F1/206—Cooling means comprising thermal management
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G06F1/32—Means for saving power
    • G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
    • G06F1/3206—Monitoring of events, devices or parameters that trigger a change in power modality
    • G06F1/3212—Monitoring battery levels, e.g. power saving mode being initiated when battery voltage goes below a certain level
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26—Power supply means, e.g. regulation thereof
    • G06F1/32—Means for saving power
    • G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
    • G06F1/3234—Power saving characterised by the action undertaken
    • G06F1/329—Power saving characterised by the action undertaken by task scheduling
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/48—Program initiating; Program switching, e.g. by interrupt
    • G06F9/4806—Task transfer initiation or dispatching
    • G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/48—Program initiating; Program switching, e.g. by interrupt
    • G06F9/4806—Task transfer initiation or dispatching
    • G06F9/4843—Task transfer initiation or dispatching by program, e.g. task dispatcher, supervisor, operating system
    • G06F9/4881—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
    • G06F9/4893—Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues taking into account power or heat criteria
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5094—Allocation of resources, e.g. of the central processing unit [CPU] where the allocation takes into account power or heat criteria
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/54—Interprogram communication
    • G06F9/547—Remote procedure calls [RPC]; Web services
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • This application relates to computer technology, and in particular, to a method, device, computer device, computer storage medium, and computer program product for resource management or resource scheduling.
  • Android Taking mobile phones as an example, the current management and control of performance, power consumption and heat are performed independently.
  • performance management and control the frequency of the central processing unit (CPU), graphics processing unit (GPU), and double data rate (DDR) memory is only adjusted based on historical load.
  • CPU central processing unit
  • GPU graphics processing unit
  • DDR double data rate
  • the CPU when the CPU load of a certain task increases, the CPU frequency increases.
  • power consumption control the execution time and power consumption of background tasks are restricted for different application types (such as instant messaging applications); or the power consumption management is graded according to the remaining power of the whole machine, such as the remaining 80% and 20 % Power corresponds to different power management levels.
  • heat sources such as SOC and display backlight
  • SOC and display backlight are limited according to whether each hot zone on the terminal device exceeds a threshold to limit the temperature rise at the cost of reducing the frequency and gear of the heat source.
  • performance, power consumption, and heat are scheduled independently according to the degree of degradation in each dimension, which is likely to cause a seesaw effect. For example, when the power is sufficient, the performance is preferentially adjusted, resulting in a rapid reduction in power and the phone continues to heat; Control power consumption and heat, resulting in reduced performance and affecting user experience.
  • the present application provides a resource scheduling method, which includes determining a current application scenario of a device, wherein the application scenario is related to user usage or an operating state of the device, and the operating scenario generally includes the following multiple dimensions: Information in any one or more dimensions: time, space, type of foreground application, and usage scenarios within the application; for this application scenario, performance, power consumption, and temperature have their own weights, based on the weights of the three and the current three Determine the resource management and control strategy based on user experience; schedule the resources of the device according to the resource management and control strategy.
  • performance, power consumption and temperature mean the respective user experience of the three.
  • the main indicator of user experience can be the standby time, then the longer the estimated standby time, the higher the user experience of power consumption; for performance, the user experience
  • the index can be the target frame rate and smoothness of the video. The higher the smoothness, the higher the user experience; for temperature, the user experience index can be the temperature of the mobile phone case. The higher the temperature of the mobile phone case, the lower the user experience .
  • the weights of performance, power consumption and temperature respectively indicate the degree of influence of the three on the overall user experience.
  • the overall user experience may be, for example, a weighted sum of the user experience of performance, power consumption, and temperature.
  • the resource scheduling strategy is determined with the overall user experience as the goal.
  • Performance, power consumption, and temperature are difficult to integrate. For example, to improve performance, power consumption will become higher and the temperature will rise. Therefore, it is difficult to find the most suitable resource scheduling strategy.
  • the method provided in this application takes into account that the importance of the three to users is different in different application scenarios. For specific application scenarios, the weights of the fusion performance, power consumption and temperature are determined to determine the appropriate resource management and control strategy, and then the resources are scheduled according to the resource management and control strategy, thus achieving a fusion decision process that comprehensively considers the three factors, effectively avoiding the three Seesaw effect brought by independent dispatcher.
  • the method provided in this application takes the weighted sum of user experience satisfaction in three dimensions of performance, power consumption and temperature as the overall user experience, and always takes the improvement of the overall user experience as the scheduling goal, effectively avoiding only one or two improvements
  • the user experience in one dimension, and the experience in another dimension decreases too much, resulting in the problem of a decrease in the overall user experience.
  • the determination of the resource management and control strategy is periodic, and the interval time period may be obtained through machine learning, or may be set by a user of the device, or preset within the device when the device is shipped.
  • resource management and control strategies can be triggered by events. Specifically, when any one or more of the performance degradation degree, the power consumption degradation degree, and the temperature degradation degree reach a preset condition, the determination of the resource management and control strategy is triggered. Among them, the degree of degradation is used to indicate the gap between the target indicator and the actual indicator. Triggering the determination (or update) of the resource management and control strategy through events can discover problems with the device in a more timely manner and provide new resource management and control strategies as soon as possible.
  • the degree of performance degradation is used to indicate the gap between the performance parameter value of the device and a target performance indicator, where the target performance indicator is one or more performance parameters that satisfy the performance experience in the application scenario
  • the value of the power consumption degradation is used to indicate the difference between the predicted available power and the lowest available power
  • the temperature degradation is used to indicate the difference between the predicted enclosure temperature and the temperature threshold.
  • the target performance indicator is determined by searching for a target performance indicator that matches the application scenario in a preset database, and the database stores multiple application scenarios and multiple target performance indicators Correspondence.
  • the database can be stored locally or on other devices.
  • the device is a mobile phone, and the database can be stored on the mobile phone or on the cloud server.
  • the target performance index can also be obtained by online estimation, or calculated and returned by the cloud server.
  • the performance parameter values of the device are collected to obtain the current performance parameter values of the device.
  • the performance parameter value at the next moment is predicted.
  • the performance parameter value of the device can be predicted according to the application scenario and the current system state of the device.
  • the system status indicates the current resource usage of the device, such as resource supply information and system load information.
  • the degree of power consumption degradation is determined by the power usage rule of the device and the remaining power of the device, where the power usage rule reflects the historical charging and power consumption conditions (or The power usage law represents the history of power increase and consumption of the device).
  • the electricity usage rule includes charging cycle, electricity consumption rule and application usage rule.
  • the charging cycle represents a historical charging interval, for example, the charging start time to the next charging start time, or the charging end time to the next charging end time, or the charging end time to the next charging start time, or the charging start time Wait until the end of the next charge.
  • the electricity consumption rule indicates a plurality of moments in history and remaining power at each moment.
  • the application usage rule reflects historical information about power consumption by applications, and may include various information, such as commonly used applications, average usage time of a single application, average usage power consumption of a single application, and / or the terminal when the application is used
  • the information such as the location of the device reflects the usage time and power consumption of the application.
  • this application provides a method for learning the user's power usage laws, including charging laws and application power consumption laws, etc., and then determining the possible degradation of power consumption according to the user's power usage laws, Instead of just presetting some fixed standby duration thresholds.
  • This method makes the determination of the degree of power consumption degradation conform to the user's power consumption habits, thereby being more accurate and effective.
  • the temperature of the housing is determined according to the current housing temperature of the device and the predicted temperature rise of each device in the device. Since each device has a different contribution to the temperature increase of the device casing, it is usually necessary to consider the weight of each device temperature when calculating the temperature of the device. The weight is used to indicate the degree of contribution to the temperature increase of the casing.
  • the “device” here can also be understood as a “hot zone”.
  • a hot zone can contain one or more devices, and the temperature can be measured with a single device as the smallest unit, or with the hot zone as the smallest unit.
  • the hot zone is usually based on the preset inside the device.
  • the temperature threshold used in the calculation of the degree of temperature degradation is determined according to the temperature of the environment in which the device is currently located. The level of the ambient temperature will affect the user's tactile temperature, so the ambient temperature is used to determine the temperature threshold, which is more suitable for the actual temperature experience of the device user, so that the calculation of the degree of temperature degradation is more accurate.
  • the resource management and control strategy includes any one or more of the following three items: a computing resource scheduling strategy, an application management and control strategy, and a peripheral management and control strategy. Resource management and control strategies are comprehensive.
  • the process of determining the resource management and control strategy utilizes machine learning or dynamic programming methods. Make full use of mathematical methods such as machine learning to speed up the process of determining resource management and control strategies.
  • the application scenarios of the device generally include a time dimension, a space dimension, and an application type dimension, such as a video watching application at home at 14:00. But the application scenario can also be more fine-grained.
  • Application scenarios may also include usage scenarios within the application.
  • the usage scenario in the application includes one or two of the following two information: the user's operation behavior on the application and the process behavior in the application. For example on WeChat Post a circle of friends or watch a video, play a team fight, etc. It is also playing games.
  • the requirements for each dimension may be different for single-player and team battle scenarios.
  • the division of fine-grained application scenarios can make the determination of user experience goals more accurate, which makes the determined resource management and control strategies more Targeted and more effective.
  • the candidate resource management and control strategy in the process of determining the resource management and control strategy, is first determined according to the performance bottleneck of the device or the dominant hot zone, and then the resource management and control strategy to be executed is determined from the candidate resource management and control strategy.
  • the candidate resource management and control strategy can increase the performance bottleneck or reduce the temperature of the dominant hot zone.
  • the factors that determine the respective weights of the performance, power consumption, and temperature include any one or more of the following: the application scenario of the device, the respective degradation degrees of the performance, power consumption, and temperature, and all
  • the action time domain for determining the performance, power consumption, and temperature degradation judgment for example, the action time domain may be in the order of milliseconds, hours, or minutes.
  • some complex calculations involved in the method may be implemented by a dedicated processor.
  • the dedicated processor can be used as a coprocessor to execute under the schedule of the main processor.
  • Special processors include but are not limited to neural network processors, machine learning processors, and so on.
  • any step in the method even if it is not complicated, can also be passed through a dedicated processor or a microprocessor or coprocessor processor.
  • the present application provides a resource scheduling method.
  • the method includes: determining a current application scenario of a device; according to the application scenario, integrating user experiences of performance, power consumption, and temperature to obtain an overall user experience; The overall user experience determines and executes resource management and control strategies.
  • the overall user experience is obtained based on the respective user experience of performance, power consumption, and temperature, and the respective weights of the three in the application scenario, for example, the weighted sum of the respective user experiences of the three as the overall user experience .
  • the weight is used to indicate the degree of influence of each factor on the user's overall experience, which may be different in different scenarios. For example, in a game scenario, users pay more attention to performance, so the weight of performance is higher than power consumption and temperature.
  • the performance experience value, the power consumption experience value, and the temperature experience value are determined according to the performance degradation degree, power consumption degradation degree, and temperature degradation degree, respectively, wherein the experience value is To indicate user experience, the degree of degradation is used to indicate the gap between the target indicator and the actual indicator. The higher the degree of degradation, the lower the corresponding experience value. Then, the overall user experience is determined according to the three experience values of performance, power consumption, and temperature, and their respective weights in the application scenario.
  • the present application provides a resource scheduling method.
  • the method includes: determining a current application scenario of a device based on a user's instruction; according to the application scenario, fusing user experiences of performance, power consumption, and temperature to obtain users Overall experience; determine and execute resource management and control strategies based on the overall user experience.
  • the method provided in this application can be in a default closed state in the device, and can be started by a user through an instruction.
  • the application scene recognition of the device may be enabled by default, but the subsequent fusion scheduling method may be enabled by a user instruction. Further, the three models of performance, power consumption and temperature can be respectively set to turn on instructions, and the user can turn on any one or two of them according to requirements.
  • the present application provides a computer device, such as a terminal device, which includes a processor and a memory.
  • the memory is used to store programs.
  • the processor is used to read and execute the program and implement the method described in any one of the foregoing aspects or implementation manners.
  • the present application provides a resource scheduling device (or resource management and control device), which includes one or more modules, and the one or more modules are used to implement any of the foregoing aspects or implementation methods. Describe the method.
  • the present application provides a computer program (or computer program product) that stores a program in the computer program (or computer program product), which is then read and executed by one or more processors to implement any of the foregoing A method described in one aspect or implementation.
  • the present application provides a computer storage medium, which may be non-volatile.
  • the computer storage medium contains a program, which is read and executed by one or more processors to implement the method described in any one of the foregoing aspects or implementation manners.
  • Figure 1 is a schematic diagram of the hierarchical division of the smart phone provided by this application.
  • 2a is a schematic diagram of a logical structure of a resource management and control device provided by this application;
  • 2b is a schematic flowchart of a resource scheduling method provided by this application.
  • FIG. 3 is a schematic diagram of another logical structure of a resource management and control device provided by this application.
  • Figure 5 is an example of the user's electricity consumption rule provided by this application.
  • FIG. 6 is a schematic diagram of the execution process of the temperature model provided by the application.
  • FIG. 7A and 7B are schematic flowcharts of the ambient temperature calculation method provided by this application.
  • FIG. 11 is a schematic diagram of a logical structure of a computer device provided by this application.
  • FIG. 12 is a schematic diagram of a logical structure of an NPU provided by this application.
  • the performance is preferentially adjusted, resulting in a rapid reduction in power and the phone continues to heat up; when the power is insufficient, it is mandatory to control power consumption and heat, resulting in reduced performance and affecting the user experience.
  • the thermal control strategy is simple and direct, and the strategy of limiting the frequency is directly adopted when the temperature is increased. With the same amount of calculation, the calculation time is lengthened after the frequency is limited, which further leads to an increase in CPU power consumption and an increase in heat, and the thermal control effect is offset.
  • the DDR frequency modulation strategy causes the DDR low frequency point to cause the CPU load to increase, and then the CPU frequency point to rise.
  • the performance bottleneck is actually memory, and the problem of low frame rate caused by the bottleneck is not solved.
  • the CPU raises the frequency band to generate heat.
  • this application provides a fusion resource scheduling method based on performance, power consumption, and temperature (ie, heating or heat).
  • the starting point of this application is the user's use scenario (equivalent to the application scenario of the terminal device), not the independent load (or degree of degradation) of performance, power consumption, and temperature, which reflects the difference in the emphasis of experience in different use scenarios. For example, game scenes have high performance experience requirements, which can reduce the experience limitation of standby time.
  • the scheduling decision of this application combines performance, power consumption and temperature, and optimizes the user's overall experience as the scheduling goal, rather than generating decision points from one or two goals, avoiding conflicts between the management and control mechanisms.
  • the method provided in the present application considers the difference between the scene and the user, and does not perform scheduling uniformly for everybody. Specifically, this application establishes an overall user experience model based on performance, power consumption and temperature. For different use scenarios, the focus of the model is different; at the same time, the load of performance, power consumption and temperature has its own Measurement models. These measurement models are built according to the individual characteristics of users or devices. The combination of the above results in a scheduling strategy based on different scenarios and with user differences.
  • the method of converged resource scheduling (or management or control) provided in this application will be described below by taking a smartphone as an example. It should be understood that the method provided by the present application can be applied not only to smart phones, but also to other types of terminal devices.
  • the terminal devices here include but are not limited to smart phones, in-vehicle devices, personal computers, artificial intelligence devices, tablets, personal digital assistants, smart wearable devices (such as smart watches or bracelets, smart glasses), and smart voice devices (such as smart Speakers, etc.), and network access equipment (such as gateways, etc.).
  • the method provided in this application can also be applied to other types of computer equipment such as servers.
  • FIG. 1 it is a schematic diagram of a logical structure of a smartphone 100 provided by this embodiment.
  • the smartphone 100 includes three layers: an application 110, an operating system 120, and hardware 130.
  • the application 110 includes various applications, such as games, WeChat, YouTube videos, browsers, and the like.
  • the hardware 130 includes a central processing unit (CPU), an application processor (AP), a graphics processor (graphics processing unit, GPU), a memory (such as DDR), an input device, and a display device (such as a touch screen is Input and display unity device), various types of sensors, etc. Some are not shown in the figure. For more hardware, refer to the description of subsequent embodiments.
  • Operating system 120 is a computer program that manages and controls computer hardware and software resources. It is system software that runs directly on the "bare metal". Various application software contained in application 110 usually needs to be in the operating system. Can only run with the support of.
  • the operating system 120 is an interface between the user and the computer, and also an interface between the hardware 130 and the application 110.
  • the functions of the operating system 120 include managing various hardware within the hardware 130, various software within the application 110, and various data resources, controlling the operation of programs, improving the human-machine interface, providing support for applications, and maximizing all resources of the computer system To a limited extent.
  • the resource management and control mechanism is the core of the operating system 120 (in this embodiment, the software program that implements this mechanism is called a resource management and control device).
  • the resource scheduling method proposed in this embodiment is a method running inside the operating system 120. This method senses the application scenarios of the upper layer, monitors the running status of the underlying resources, and combines the three factors of performance, power consumption, and temperature to determine the resource scheduling strategy to enhance the overall experience of the smartphone 100.
  • the resource scheduling method provided in this embodiment runs inside the operating system 120, and mainly includes three modules: a scene awareness module 121, a resource monitoring and decision module 122, and an execution module 123. These three modules respectively perform the three steps S201-S203 shown in FIG. 2b.
  • Application scenarios usually include information in one or more of the following dimensions: time, space, type of foreground application, and usage scenarios within the application.
  • time time
  • space space
  • type of foreground application usage scenarios within the application.
  • usage scenarios within the application For example, the scenario of "watching video at home” includes two dimensions: space and the type of foreground application.
  • the application scenario of the mobile phone can be understood as the usage scenario of the user using the mobile phone, such as a meeting scene, a sleeping scene, a work scene, a home scene, a noise scene, a driving scene, a running scene, etc.
  • the application scenarios of mobile phones can be single-dimensional or multi-dimensional, such as users watching videos in a noisy environment. Further, if the user is using an application, they can continue to subdivide different scenarios within the application, such as watching 1080p or 720p video, or playing a team battle in a game, etc.
  • the action or movement state of the mobile phone can be classified into the "space" dimension.
  • the foreground application type of the running scene may be a health application or a music application, and the sensing device on the mobile phone senses that the mobile phone is in regular movement, that is, the "running scene" includes the type and space of the foreground application Dimensional information.
  • S202 Determine a resource management and control strategy based on the weights of performance, power consumption, and temperature in the application scenario.
  • the trigger condition for determining the resource management and control strategy may be that any one or more of the degree of performance degradation, the degree of power consumption degradation, and the degree of temperature degradation reach a preset condition.
  • the degree of performance degradation is used to indicate the gap between the performance parameter value of the device and a target performance index, where the target performance index is the value of one or more performance parameters that satisfy the performance experience in the application scenario.
  • the degree of power consumption degradation is used to indicate the difference between the predicted available power and the lowest available power.
  • the degree of temperature degradation is used to indicate the difference between the predicted case temperature and the temperature threshold.
  • the "gap” between the two items in this embodiment refers to the difference between the two items.
  • the difference between these two items the ratio between these two items, or the value obtained by adding other factors to the difference or the ratio; or it is not a specific value , Is a function that shows the trend.
  • the development of performance parameter values is contrary to the target performance index. For example, performance will decline, but the target performance index is a higher value. In other words, as long as it can characterize the degree of degradation, this implementation does not limit the specific implementation.
  • Performance, power consumption and temperature here can be understood as performance experience, standby time (ie power consumption experience), and case temperature (ie temperature experience).
  • the target performance index first determines the target performance index according to the application scenario determined in the previous step. Then, calculate the gap between the target performance index and the performance parameter value currently configured on the mobile phone to determine the degree of performance degradation, and determine whether to generate a performance degradation alarm according to the degree of performance degradation.
  • a degree of power consumption degradation is determined according to the application scenario, power usage habits, and current remaining power of the terminal device, and whether to generate a power consumption degradation alarm is determined according to the degree of power consumption degradation.
  • the temperature of the casing of the terminal device is predicted based on the collected temperature of the hot zone, the degree of temperature degradation is determined according to the temperature of the casing, and whether a temperature rise alarm is generated is determined according to the degree of temperature degradation.
  • the user may set a resource management and control period to periodically trigger the policy determination process.
  • S203 Schedule resources of the device according to the resource management and control strategy.
  • the resource management and control strategy includes any one or more of the following three items: computing resource scheduling strategy, application management and control strategy, and peripheral management and control strategy.
  • the resource management and control strategy determined in the previous step is a resource management and control strategy that satisfies the optimal user experience in the application scenario.
  • the user experience is obtained by satisfying the performance, standby time, and temperature, and the weights of the three in the application scenario through a certain mathematical transformation.
  • the process of determining resource management and control strategies can be understood as solving the mathematical process of optimal user experience, such as machine learning or dynamic programming.
  • "optimal" is not necessarily understood as the theoretical or mathematical maximum.
  • the user experience is the target experience. This value can be set by the user, or although this value is not the mathematical maximum, it may be the most appropriate value in the current scenario.
  • the mobile phone can perceive the usage scenario where the user is located, determine the target resource demand (equivalent to the target performance index) according to the usage scenario, and calculate the gap between the performance demand and the current resource configuration
  • the target resource demand equivalent to the target performance index
  • the gap between the performance demand and the current resource configuration To obtain performance degradation data, combined with the degradation data generated in both power consumption and heat, scheduling decisions are made based on the weights of the three aspects of performance, power consumption and heat in the current usage scenario and issued to ensure the performance of the foreground at the same time , The standby time and the temperature of the mobile phone case are in a steady state.
  • the method provided in this application combines the performance, power consumption and heat according to the weight of the current scenario under the premise of identifying the user's usage scenario, and optimizes the combined user experience of the three as the The goal of scheduling, instead of generating decision points from one or two goals, avoids conflicts between the management and control mechanisms, and reflects the focus of the three in different usage scenarios, thereby generating more effective resource control strategies.
  • FIG. 3 is a schematic diagram of further module division of the resource management and control apparatus provided in this embodiment, in particular, a more detailed description of the resource monitoring and decision module.
  • the device includes a scene perception module 210 (equivalent to the scene perception module 121), an execution module 260 (equivalent to the execution module 123), and 220-250 four modules.
  • the scene perception module 210 is used to identify the application scenarios of the terminal.
  • the application scenario is identified by one or more dimensions of information. For example, user usage scenarios (home / office, day / night, workday / rest day), application types (such as video, games, etc.), fine-grained usage scenarios within applications (such as watching 1080p HD video), etc.
  • the performance model 220 is used to predict whether performance will deteriorate or the degree of performance degradation. Specifically, the performance model 220 is used to perceive the target performance goals in different application scenarios, and combine with the current system state (including resource supply status and system load status) to predict and judge whether the user performance dimension experience will deteriorate and the degree of deterioration. Further, the performance model 220 will analyze and identify the main bottleneck that causes performance degradation based on the PMU (performance monitor unit) information reported by the chip.
  • PMU performance monitor unit
  • the performance model prediction is usually in the millisecond level.
  • the power consumption model 230 is used to predict the degree of power consumption degradation.
  • the degree of power consumption degradation can be expressed by the length of standby time or the degree of power degradation.
  • the prediction of the power consumption model is usually at the hour level.
  • the thermal control model 240 is used to predict the degree of temperature degradation. In some embodiments, according to the thermal steady state algorithm and the shell temperature prediction algorithm, it is determined whether the temperature of the mobile phone exceeds a threshold for a period of time (for example, 8 minutes), causing severe fever, and if it exists, it will be issued. Further, the thermal model will identify the dominant hot zone that causes heat generation based on the temperature of the hot zone on the mobile phone and the degree to which the hot zone affects the final case temperature.
  • the prediction of the thermal control model is usually at the minute level.
  • the decision module 250 is used to determine a resource management and control strategy based on the weights of performance, power consumption, and temperature in the current application scenario.
  • the fusion scheduling decision is triggered based on the alarm information from the performance model, power consumption model, and thermal model, taking into account the needs of the above three dimensions, the status quo, and the weights of the three dimensions in the current application scenario to dynamically determine The ultimate solution for the best comprehensive user experience.
  • the execution module 260 is used to implement resource management and control according to the determined resource management and control strategy.
  • Resource management and control usually involves multiple management and control measures for multiple resources, so the scheduling management and control module can be understood as a combination of multiple functional units distributed in different places of the system.
  • Resource management includes scheduling of real-time resources, background processes, and peripherals.
  • Scene perception is the synthesis of multiple recognition methods, outputting the user's specific use scenarios, such as the current user watching hard-decoded video with 1080p resolution at home.
  • Scene perception includes three aspects: user space-time perception, application type perception, and in-app scene recognition.
  • the user's space-time perception can include time dimension and space dimension.
  • Time dimensions such as holidays / workdays, day / night. Differentiate the time by obtaining the current system time or querying from the system calendar.
  • Spatial dimensions such as travel / tourism, home / office. Clustering is based on the collected GPS location information, and the home and office locations are identified according to the most frequently occurring geographic location points at different times; combined with the current time to determine whether to travel or travel.
  • Application type awareness is to perceive what application the user is using in the foreground. You can preset a list of multiple application types in the mobile phone system, for example, the top 10,000 domestic user types, and then query the list according to the current foreground application to determine the corresponding application type of the foreground application, such as WeChat (WeChat) Instant messaging applications.
  • the mobile phone may also send the current application identification to the cloud server through the network, and the cloud server recognizes and returns the identification result to the mobile phone.
  • In-app scene recognition is to identify more fine-grained scenes within the application.
  • the dimension of the in-application scenario includes the user's operation behavior on the application (referred to as user behavior) and the behavior of the process within the application (referred to as process behavior).
  • the user's behavior such as playing videos with different definitions and conducting video chats, belongs to the user's behavior, and the game's start-up, magnification, and hero selection belong to the in-app process.
  • User behavior can be obtained by reporting various events and performing metadata identification in the data stream.
  • Process behavior can be identified by the application's drawing information or SDK pre-embedded in the application to actively report to the system.
  • the performance model is mainly used to identify performance targets in different scenarios, and combined with the current system state (including resource supply and system load) to predict and judge whether the user's performance dimension experience will deteriorate (millisecond level), and the degree of degradation . Further, the performance model will analyze and identify the main bottlenecks that cause performance degradation based on the current usage scenario and system status.
  • FIG. 3 is a schematic flowchart of a method of scene recognition and determining a performance degradation alarm. The method is roughly divided into the following three parts a), b) and c).
  • the process of implementing scene recognition in various ways has been introduced in (1) above, and will not be repeated here, and reference can be made to FIG. 4.
  • a target performance index matching the scene is searched in the database 301 set in advance (S320).
  • the database 301 stores correspondences between various application scenarios and various target performance indicators.
  • the database can be generated offline and pre-configured.
  • Ti includes: target frame rate, frame loss rate, low frame rate ratio (the percentage of frames below the target frame rate in the total number of frames), and fluency (frame rate fluctuations), etc.
  • the scene division and experience goals in the database may be formulated through manual participation.
  • the scene division and target performance indicators in the database are discovered by clustering the reported big data, for example, the video playback scene is subdivided into different frame rates (24/25/30 frames) And soft and hard decoding scenarios, respectively corresponding to different experience goals.
  • the scene sensing module 210 may be used to identify a scene S i, then the performance model 220 queries database 301 to obtain the target scene performance indicator S i T i.
  • the scene perception module 210 determines that the 30-frame scene is hard-decoded through the decoding method of the video stream data and the frame header information.
  • the frame loss rate is less than 30%, the low frame rate is not more than 20%, and the fluency is within 0.8.
  • the frame rate change of the next sampling period (millisecond or second level) is predicted and the performance index of other dimensions is obtained through calculation (S330).
  • the system status includes resource supply status and system load status.
  • Performance indicators in other dimensions such as frame loss rate, low frame rate ratio, and fluency, can be calculated based on the current frame rate change and the corresponding calculation formula. Calculated as follows:
  • the specific calculation method is:
  • r represents a certain dimension of resources
  • R represents the collection of all resources
  • r cur represents the current resource supply status
  • r within_scene represents the demand of resources in this application scenario
  • ⁇ app represents the weight coefficient of different applications on different resources
  • ⁇ overload represents the weight coefficient corresponding to the system load.
  • ⁇ app and ⁇ overload can be obtained by the following methods: the first one, which is preset in the mobile phone after offline analysis; the second one, online analysis, such as statistics on the usage of various resources of the application at runtime, and the analysis of this application for different The focus / requirement of resources, based on which to learn the weight of each dimension.
  • the performance index under the current state of the system can be obtained, and it can also be obtained through data collection.
  • the frame rate for a period of time is collected and obtained, and the performance indexes of other dimensions such as average frame rate, frame drop rate, low frame rate ratio, fluency, etc. are obtained through collection or calculation (S340).
  • predicted and collected data can be used either, or the data obtained by these two methods can be used after some processing, such as taking the average of the two.
  • the gap between the current performance and the target performance that is, the degree of performance degradation can be calculated (S350).
  • the actual feedback performance index can be obtained through the prediction method mentioned above, or can also be obtained through real-time collection.
  • ⁇ i is the degradation weight of the i-th index in different scenarios.
  • ⁇ i 1.
  • the weight corresponding to the deterioration of the average frame rate in the game scene is greater than the weight of the average frame rate in the video scene.
  • T target i represents the target value of the i-th index.
  • the target performance index of the average frame rate of the game scene in the game is 57 frames.
  • T out i represents the actual value of the i-th index.
  • the other steps can be performed by the performance model 220.
  • the power consumption model is used to learn the user's power usage habits and evaluate the impact of the current application scenario and current application on the standby time or available power. For certain users, the two variables of standby time and available power can be converted between each other, so the following will only use the available power as an example.
  • the specific implementation of the power consumption model generally includes the following three parts a), b) and c).
  • the information collected and recorded includes but is not limited to the application used by the user, the duration of use, the charging start and end times, the charging location, and the average power consumption of the application.
  • Persistent storage the storage form is a file or database.
  • the permanently saved power usage information in step a) is used as input, and the user's power usage rules are learned through machine learning or statistical methods.
  • the power usage rules include charging cycles, power usage laws, and application usage laws.
  • the charging start time and the charging end time are used as inputs to learn the charging cycle.
  • Different users may have different charging cycles, for example, some people charge once a day, and some people charge once every two days.
  • Use the applications used by the user, the duration of the application, the average power consumption of the application, the user's location, etc. as input to learn the commonly used applications of the user in different locations and time periods, as well as information such as the average duration and average power consumption of the application .
  • step b) Taking the charging cycle and electricity consumption law learned in step b) as input, through time period division and statistical methods, each time period and the corresponding minimum available power amount are output.
  • the time period is based on the charging cycle rather than a fixed day. Different users have different charging cycles. For example, some people charge once a day, that is, the charging cycle duration is 24 hours; some people charge once every two days, that is, the charging cycle duration is 48 hours.
  • the method for dividing the time period may be a fixed time slice, for example, every 3 hours as a time period.
  • the division of time periods may be based on semantics, such as morning, morning, noon, afternoon, evening, evening, and midnight.
  • the time segment can use a dynamic division method.
  • An example of the dynamic division method of the time period is as follows: First, the lower limit number of the preset time period is used, for example, 3, and the charging cycle is divided equally. Determine the duration of each time period. If the duration exceeds a threshold, such as 3 hours, then interpolate in this time period to ensure that the duration of each time period does not exceed the duration threshold. Evaluate the power change or standby time change of each time period. If the power change or standby time change of a time period exceeds the threshold, for example, the minimum available power in the previous period is 80%, and the minimum available power in the latter period is 20% The power change is 60%, and the threshold exceeds 20%, then interpolation is performed during this time period to ensure that the power change in each time period does not exceed the power threshold.
  • a threshold such as 3 hours
  • FIG. 5 is a schematic diagram of the divided time periods, in which the minimum available power calculation method for each time period is as follows:
  • ⁇ ( ⁇ ) represents an elastic function, that is, allows the lowest available power to fluctuate within the value range represented by ⁇ (std, location, scene), which is based on historical statistics The standard deviation std, the current location and the scene of the current application scene as variables.
  • This step is executed after step c) is completed. This step is executed periodically or triggered when the application used by the user changes.
  • the input of this step is the current power, the application usage rules learned in step b) and the time period division information calculated in step c), and the output is the power consumption deterioration assessment and the degree of deterioration.
  • the application usage rule and the current foreground application predict the usage time of the foreground application and the estimated power consumption ⁇ E in the next time period, and then according to the minimum available power to be met in the next time period Evaluate whether power consumption has deteriorated. If the current remaining power E cur minus the estimated power consumption ⁇ E is lower than the minimum available power to be met in the next time period Then, the power consumption will deteriorate, and the alarm will be sent to the fusion scheduling module for decision.
  • the evaluation of the deterioration degree ⁇ can be calculated by the following publicity:
  • Thermal control model (corresponding to thermal control model 240)
  • the thermal control model will identify the dominant hot zone that causes heat based on the hot zone temperature and the extent of the hot zone's influence on the final temperature.
  • the implementation of the thermal control model is mainly divided into the following parts a) -e).
  • the layout of the mobile phone including the structural layout of the mobile phone and the material of the shell of the mobile phone, will affect the heat dissipation capacity of the mobile phone itself.
  • function approximation methods such as the least square method, offline learning of the mobile phone's heat dissipation capacity curve under various usage scenarios, and pre-configuration before the device leaves the factory.
  • the shell temperature is the shell temperature of the phone.
  • the temperature of the housing is determined according to the current housing temperature of the mobile phone, the predicted temperature rise of each device in the mobile phone, and the weight of the temperature of each device.
  • the predicted temperature rise value of the device is determined according to the temperature rise function of the device and the current resource configuration of the device, where the temperature rise function may be preset in the mobile phone.
  • the specific calculation formula is:
  • ⁇ i is the weight of each hot zone temperature, which is related to the layout of hardware devices and so on.
  • T now is the temperature value sampled in the hot zone at the current moment.
  • the temperature rise temperature of each hot zone at time T in the future can be measured by F i (R i , t).
  • R i is the current system resource allocation (or supply) situation
  • t is the duration.
  • F i is the temperature rise function of the i-th hot zone device (such as CPU or GPU or DDR), which is related to the hardware configuration and layout of the product. It is generally the result of linear fitting of multiple factors and is a fixed function.
  • DDR stands for double rate
  • DDR SDRAM stands for double rate synchronous dynamic random access memory. It is used in the art as DDR.
  • SDRAM is an abbreviation for synchronous dynamic random access memory, namely synchronous dynamic random access memory.
  • a “hot zone” may contain only one device or multiple devices of the same type or different types.
  • ⁇ j is the ratio of the hot zone j to the shell temperature
  • ⁇ j * F j (R j , t) is the increase in the shell temperature generated by the hot zone j
  • ⁇ i * F i (R i , t) is The rise in shell temperature in all hot zones.
  • Identifying the dominant hot zone can provide input for the scheduling decision of the decision module 250.
  • the input for the calculation of ambient temperature includes: temperature data collected by the NTC node, power consumption data of each device, and ambient temperature data of the thermostat for simulation training, and the fitting parameters of the ambient temperature model are calculated.
  • the process is shown in Figure 7A. First, the current system state and the stored state history table collected by each NTC node (S701), and then determine whether the ambient temperature needs to be updated (S702 and S703). If the interval time from the last ambient temperature output does not exceed the interval threshold (S702) , It will not be updated, otherwise it will continue to judge whether it is in the charging state (S703), if it is in the charging state, it will not be updated.
  • NTC thermostat temperature sensor.
  • the NTC node refers to a temperature collection sensor, which is used to collect the temperature of each hot zone.
  • the calculation of the ambient temperature has a certain period.
  • the time from the last calculation exceeds the threshold and the current mobile phone is not in the charging state, the latest ambient temperature calculation is performed, and the calculation result is added to the ambient temperature history table.
  • the Kalman filter-like calculation method is used for the calculation of the ambient temperature.
  • a linear time-invariant reduced-order model (LTI-ROM) can be used to predict the ambient temperature.
  • the inputs to the model are multiple power consumption P1-P5 and the collected temperature values.
  • the output of the LTI-ROM model is the predicted value of the ambient temperature at the next moment.
  • the predicted value is input into the correction model.
  • the corrected model combines the actual value of the ambient temperature to correct the predicted value. When the predicted value and the actual value are consistent, the ambient temperature is output. , When inconsistent, reverse adjust the parameters of the LTI-ROM model.
  • the judgment conditions of the temperature rise alarm can include the following two: the maximum temperature in a short period cannot exceed the threshold, and the temperature fluctuation range cannot exceed the threshold during this time.
  • T The temperature of the whole machine in time period T is described by H (T), T thresold is the upper temperature threshold, and ⁇ is the temperature fluctuation threshold.
  • T the temperature of the whole machine in time period T is described by H (T)
  • T thresold is the upper temperature threshold
  • ⁇ the temperature fluctuation threshold.
  • the significance of the above formula is that the highest temperature in the time period T is less than T thresold , and the difference between the highest temperature and the lowest temperature is less than ⁇ . In other embodiments, “less than” may also be "less than or equal to”.
  • the temperature threshold T thresold and / or ⁇ can be adjusted according to the ambient temperature. For example, if the ambient temperature is very low, then T thresold can be set higher, because at this time people have a higher tolerance for the temperature of the phone case; on the contrary, if the ambient temperature is high, then T thresold can be set lower, because at this time people The temperature tolerance of the mobile phone case is relatively low.
  • the above two conditions are not indispensable, and any one of them may be selected as the judgment condition of the temperature rise alarm.
  • the composition of strategy a can be ⁇ CPU / GPU / DDR gear, peripheral control gear, application management strategy ⁇ .
  • the overall user experience R is a weighted value of the following factors:
  • S includes the following dimensions: performance experience goal satisfaction degree (S 1 ), standby duration (S 2 ), and shell temperature experience (S 3 ).
  • S 1 performance experience goal satisfaction degree
  • S 2 standby duration
  • S 3 shell temperature experience
  • the overall experience R is related to three factors: performance, power consumption, and case temperature.
  • ⁇ i is the weight of each factor, and the weight is different in different terminal application scenarios (or different users). For example, the weight of satisfaction of performance experience goals in game scenarios is higher than reading e-book scenarios.
  • the goal of the search is to find the strategy a that maximizes R, ie max a ⁇ A R.
  • the selection process of fusion scheduling decision can be abstracted into a reinforcement learning or Markov decision process.
  • reinforcement learning Take reinforcement learning as an example,
  • the environment X represents the current state, that is, ⁇ use scenario, performance / power consumption / thermal degradation degree ⁇ .
  • A represents the currently selectable strategies, such as the combination of resource supply adjustment, peripheral control strategy and application management strategy.
  • P is the state transition probability matrix learned in offline environment.
  • R is the reward of making A strategy in the X state, that is, the optimization / deterioration change of the user's overall experience;
  • fusion scheduling decisions will learn and update subsequent strategy choices based on runtime system and user feedback.
  • the feedback involved includes:
  • Performance dimension frame loss, freeze, peripheral active adjustment feedback (adjusting backlight, volume), subsequent applications, etc .;
  • Power consumption dimension the number of times the battery is exhausted, whether the power consumption exceeds a certain period of time, etc .
  • Temperature dimension the number of times the battery heats up, whether the battery temperature has effectively decreased, etc.
  • the following combines an example of a game scene and the fusion scheduling method provided by the present application in the foregoing method.
  • the method is mainly divided into four perception modules: scene perception, performance model, power consumption model, and thermal control model. It integrates a decision module with a scheduling center (including decision library) and an execution module that interfaces with an external scheduling mechanism.
  • Scene awareness is responsible for identifying fine-grained usage scenarios.
  • the performance model is a millisecond-level perception module, which is responsible for determining the target performance requirements according to the usage scenario, as well as determining the current performance indicators, and then determining the degree of performance degradation and performance bottlenecks based on the two.
  • the power consumption model is a perception module at the hour level, which is responsible for judging the degree of deterioration of the power consumption (that is, the standby time) and outputting an early warning.
  • the thermal control model is a perception module at the minute level, which is used to judge the degree of temperature rise and dominate the hot zone.
  • the fusion scheduling center receives the input of the three perception modules of the performance model, power consumption model, and thermal control model, and then makes the scheduling strategy with the best overall performance and delivers it to the execution module.
  • the scheduling strategy may include computing resource scheduling, background application management and control, and peripheral control.
  • the decision library is used to store some preset scheduling strategies for the fusion scheduling center to search for to obtain the scheduling strategy with the best overall performance.
  • Perceive the user's usage scenarios that is, mobile phone application scenarios
  • the basic input of the performance model used to identify the performance requirements of the user's foreground.
  • the user's usage scenario includes several dimensions, which are provided by the following data sources.
  • the user's time dimension and spatial dimension including the current time and the type of location of the current user.
  • the time dimension can be queried through the system calendar, such as holidays and working days.
  • the spatial dimension includes home / office / others, which can be obtained through location aggregation services.
  • some mobile phones provide context intelligent services, which can be obtained by querying the context intelligent service interface.
  • Application type dimension such as instant messaging application, mail application or game application.
  • the application type generally represents the core use of the application.
  • the type of WeChat is an instant messaging application.
  • the source of the data may be to query the application type database (database is set on the end side or the cloud side), or dynamically identify on the end side, or query from the cloud.
  • the in-app scene is the video playback scene.
  • the scene can be identified according to the application behavior events reported in the framework (Framework) and the foreground application related information.
  • the video playback scene can be identified based on the full-screen events and video playback events (or video type applications and vocalization events) reported by the dot.
  • ActivityManagerService reports the activity com.tencent.mm/com.tencent.mm.plugin.sns.ui.SnsTimeLineUI in the foreground, it can be known that the user is currently using the friend circle function of the WeChat application.
  • in-app scene parameters such as Youku's in-view video to distinguish between full-screen / non-full-screen, soft-decoded video / hard-decoded video, 360p, 720p / 1080p and other different resolutions, and whether there are user subtitles displayed.
  • This part of the data can be obtained by dotting in the framework, or the application that integrates a specific SDK actively sends related parameters.
  • the user's perception of usage scenarios is encapsulated into fine-grained usage scenarios represented by the scene data type and passed to the performance model and power consumption model as inputs to these two models.
  • the scene information output by the user's scene perception is: ⁇ holiday, night, home, G game, team battle , 60 frames>.
  • G refers to the name of the game application
  • Team Battle refers to a multiplayer war scene in the game.
  • the performance model is used to discover the degree of performance degradation, send out performance degradation awareness information and performance bottlenecks, and notify the fusion scheduling center.
  • performance experience indicators usually include the following dimensions: average frame rate, low frame rate ratio, and jitter rate information.
  • the performance experience index may be the same as or different from the game scenario, which is not limited in this embodiment.
  • the performance model is executed in the order of 1-4 in the following order, and finally the performance degradation degree and performance bottleneck are output.
  • Scene ⁇ holiday, night, home, G game, team fight, 60 frames>, query in the database to obtain the performance index corresponding to the current scene.
  • performance indicators can be distinguished according to the classification of mobile phones. For example, latest machines, mid-range machines, and low-end machines correspond to performance targets A, B, and C, respectively.
  • the three mobile phone levels can be implemented as three tables in the database, and each table contains the correspondence between the mobile phone level scenarios and performance indicators; it can also be a table, distinguished by the mobile phone level field.
  • the operating environment is the main machine with a resolution of 1080p
  • the performance target A corresponding to the current Scene is determined to be ⁇ average frame rate 57, low frame rate accounts for 0.5%, and jitter rate 0.2% >.
  • the corresponding performance target B is ⁇ average frame rate 56, low frame rate accounts for 1%, and jitter rate 3%>. If it is a low-end machine, the corresponding performance target C is ⁇ average frame rate 55, low frame rate 1%, jitter rate 1%>
  • the performance model can predict changes in performance indicators in the short term based on the current system load and SOC resource supply.
  • Surfaceflinger is a component of the Android system framework layer.
  • the full name of SOC is system on chip, which can be considered as a chipset and integrates each core of the processor. In this embodiment, it includes 4 large cores and 4 small cores. In this implementation, GPU and DDR are also on the SOC.
  • performance degradation can be found. Quantify this performance degradation, and generate different levels of performance degradation alerts.
  • the performance bottleneck discovery mainly depends on the performance monitoring unit (performance monitoring unit, PMU) reported event on the SOC for identification. It is an existing mature solution and will not be repeated here.
  • PMU performance monitoring unit
  • the quantization method of performance degradation is the weighting of three experience indicators.
  • For the quantization method refer to the aforementioned formula (3).
  • the weights of the deterioration levels of the three indicators are: average frame rate 0.5, low frame rate 0.3, and jitter rate 0.2, corresponding to ⁇ i in formula (3).
  • the average frame rate has the largest weight in the G-game team battle scenario; the low frame rate comes second; the weight of the jitter rate is the smallest.
  • the quantification of the deterioration degree of each dimension is defined as the percentage difference between the current performance index T out i and the performance target T target i .
  • the average frame rate target in performance target A is 57 frames
  • the current average frame rate is 50 frames
  • the target value is 0.5
  • the current value is 0.3
  • the current value is better than the target value, so the degradation value is a negative number, and the degradation value can be forcedly set to 0 at this time.
  • Different ⁇ T values correspond to different levels of performance degradation alarms, and the performance degradation alarms are reported to the fusion scheduling center. Specifically, 0.16 can be directly reported, or 0.16 can be converted into a level of performance degradation alarm, for example, the corresponding level is 0.3, and 0.3 is reported to the fusion scheduling center.
  • the power consumption model also needs to consider the current usage scenario in this embodiment, because the power consumption in different usage scenarios may be different, such as the game scenario in this implementation, the power consumption will be faster than the chat scenario.
  • the operation process of the power consumption model mainly includes the following two steps.
  • the power consumption model collects power usage information in the background.
  • the power usage information is divided into two parts: charging habits and power consumption habits.
  • n 7, that is, the last 7 charge records as a window, and calculate the variance of the charging start time period in the window. For example, if the charging times in the past 7 times are 9 o'clock, 10 o'clock, 9 o'clock, 9:30, 8:30, 9 o'clock, and 10:30, the corresponding variance is 0.41.
  • the variance of the charging record is less than a certain threshold, it is considered that the user's charging habits are stable, and the time prediction of the user from the next charging under the current space-time type can be obtained. If the variance of the charging record is relatively large, the default charging duration (12 hours) is used as the prediction to calculate the time interval from the next charging.
  • each time slice is 3 hours. Count the applications that users commonly use in the current time slice and the power consumption per unit time of the application, and weight to obtain the power consumption overhead of the user in the time slice as the user's power consumption habits.
  • the power consumption deterioration evaluation is performed.
  • the way to evaluate power consumption degradation is to calculate the degree of impact of the current application on the standby duration at the end of this time slice.
  • the user starts to play the G game.
  • the user's charging habits show that the user may start charging at 21:00, and the current power reserve is 3000 mA.
  • the power consumption model needs to reserve 20% of the power to avoid entering the power saving mode, so the current distributable power is 2800mA.
  • the two time slices of 15: 00-18: 00 and 18: 00-21: 00, the user's historical power consumption ratio is 3: 7, then the two time slices are allocated 840mA and 1960mA respectively.
  • the deterioration of power consumption triggers a power consumption alarm, and sends the deterioration alarm to the fusion decision module.
  • the mobile phone Before leaving the factory, the mobile phone will preset two parts of information related only to the device model and the layout of the whole device:
  • the temperature rise function F i (R i , t) of the device in the hot zone which means that the temperature rise generated by the device after the gear is at R i and duration t;
  • the ambient temperature is estimated by means of Kalman filtering when it is not in a charging state.
  • Kalman filtering For the specific method, refer to FIG. 7A and FIG. 7B described above.
  • the temperature of the hot zone is collected regularly, and the predicted value of the shell temperature is calculated according to the collected hot zone temperature and the temperature rise prediction function of the hot zone. Then generate a temperature rise warning based on the prediction of the shell temperature and the ambient temperature.
  • the current ambient temperature is 26 degrees.
  • the CPU runs at high load and high frequency. Through the temperature collection of each hot zone and the prediction of the hot zone temperature, it is found that the current shell temperature of 38 degrees will exceed 42 degrees in 5 minutes, and the highest temperature in the short term (42 degrees) exceeds the threshold (assuming the threshold is 40 degrees), Generate a temperature rise warning.
  • the quantified temperature degradation result which converts the warning information like ⁇ short-term temperature rise, current 38 degrees, more than 42 degrees in 5 minutes> into a temperature degradation value.
  • This conversion can be carried out according to formulas or according to rules.
  • the temperature is divided into several gears: 38,42,45,48,53, the temperature rise in 5 minutes exceeds the corresponding gear corresponding to the temperature degradation value 0.2,0.3,0.5,0.8,1.2, the temperature in 10 minutes If the rise exceeds the corresponding gear, the corresponding temperature degradation value is 0.1, 0.2, 0.3, 0.4, 0.6.
  • the temperature degradation value output is 0.3.
  • the discovery of the main thermal conduction zone was carried out during the calculation of the shell temperature.
  • the hot zone that contributed the most to the increase in shell temperature was used as the main thermal conduction zone (for example, GPU) and sent to the fusion dispatch center for decision-making.
  • the weight coefficient ⁇ of each user experience dimension is different for different application scenarios and different users, and can be determined by the following factors:
  • ⁇ (.) Is a weight calculation formula, and the weight coefficients of the three dimensions are calculated according to multiple factors, including:
  • Scene is a terminal application scene, such as game scene, video scene, reading scene, etc.
  • the importance of each dimension is different in different scenarios.
  • the weight of the performance dimension is higher than the weight of the power consumption and thermal dimensions; in this embodiment: Frame>;
  • time_domain is the time domain of the experience dimension, such as seconds, minutes, and hours. The smaller the time domain of the experience dimension, the greater the relative weight;
  • Deterioration is the degree of deterioration of this dimension. The more severe the degree of deterioration, the greater the weight of this dimension.
  • the final calculated ⁇ performance dimension weight, power consumption dimension is ⁇ 0.5,0.3,0.2>.
  • the calculation formula of ⁇ can be preset data to be queried or can be calculated in real time according to the formula.
  • the user experience in the performance dimension measured according to the front-end user experience and the back-end user experience.
  • the front-end user experience is determined by the target frame rate, frame loss rate, low frame rate ratio and fluency
  • the back-end user experience is measured by the proportion of hot-starts used by background applications in the subsequent use.
  • the quantification method of the user experience in the performance dimension is the satisfaction of the three indicators of average frame rate, low frame rate ratio, and frame rate jitter.
  • the calculation method is 1 minus the degree of performance degradation 1- ⁇ T, that is, the more The smaller the better the user experience, the maximum is 1.
  • User experience in the power consumption dimension measured according to standby time.
  • the standby duration does not reach the set upper limit, the greater the standby duration, the better the user experience in the power consumption dimension; after the upper limit is reached, the increase in the standby duration does not affect the user experience in the power consumption dimension.
  • the upper limit of the standby time can be 1.2 times the user's charging cycle.
  • the user experience quantification method in the power consumption dimension is the ratio of the predicted standby duration to the upper limit of the standby duration, and the maximum is 1.
  • the user experience of the temperature rise dimension when the subsequent temperature does not exceed the threshold (that is, the user feels that the phone is not hot), the user experience of the temperature rise dimension is measured according to the maximum value and variance of the temperature rise speed, the smaller the maximum value and the greater the variance Smaller is the best temperature rise experience (control the temperature rise speed to rise at the most gentle speed).
  • the thermal dimension of user experience is quantified by 1 minus the degree of temperature rise and degradation. The maximum is 1. For the quantification method of the temperature rise and deterioration degree, please refer to "Process 4. Thermal Control Model".
  • FIG. 10 is an example of a process for determining an optimal scheduling strategy. As shown in the figure, the method of finding the optimal strategy in the available control strategy space can be divided into the following steps.
  • S901. Determine the direction of policy search with the help of performance bottlenecks and dominant hot spots, and select candidate resource scheduling policy sets in a targeted manner.
  • the performance bottleneck and the dominant hot zone are provided by the performance model and the thermal control model, respectively, which can help reduce the search space of the strategy and reduce the number of candidate resource scheduling strategies.
  • the performance bottleneck of the G game is DDR.
  • DDR resources should be increased to improve the performance experience;
  • the dominant hot zone is the GPU, the frequency of the GPU can be targeted to achieve a rapid Control temperature rise.
  • the dominant hot zone and performance bottlenecks are large nuclear resources, then the frequency of large nuclear resources rises and falls are possible search directions, and the overall experience of the users behind To decide.
  • resource power consumption table information can also be used as an input to filter resource scheduling strategies.
  • the resource power consumption table information is provided by the power consumption model.
  • the resource power consumption table includes various resources and their corresponding power consumption information, for example, at what level of computing resources, such as CPU and GPU, and how much power is consumed per unit time; at what level of peripherals, such as screen and sound, etc. , How much power is consumed per unit time; how much power is consumed by different applications in the background per unit time.
  • FIG. 11 is a schematic diagram of a device structure of a smart phone.
  • Figure 1 is a layered schematic diagram of the smartphone system architecture, while Figure 11 depicts the structure of the phone from the hardware perspective as much as possible.
  • FIG. 11 for the hardware included in the hardware 130 in FIG. 1 and the connection relationship thereof.
  • the operating system 120 and the application 110 in FIG. 1 are stored in the memory 870 shown in FIG. 11 in the form of software programs.
  • the mobile phone includes a communication module 810, a sensor 820, a user input module 830, an output module 840, a processor 850, an audio and video input module 860, a memory 870, and a power supply 880. Further, the mobile phone provided in this embodiment may further include a neural-network processing unit (NPU) 890.
  • NPU neural-network processing unit
  • the communication module 810 may include at least one module that enables communication between the computer system and a communication system or other computer systems.
  • the communication module 810 may include one or more of a wired network interface, a broadcast receiving module, a mobile communication module, a wireless Internet module, a local area communication module, and a location (or positioning) information module. These multiple modules all have multiple implementations in the prior art, and this application does not describe them one by one.
  • the sensor 820 may sense the current state of the system, such as the open / closed state, position, whether there is contact with the user, direction, and acceleration / deceleration, and the sensor 820 may generate a sensing signal for controlling the operation of the system.
  • the user input module 830 is configured to receive input digital information, character information, or contact touch operation / contactless gestures, and receive signal input related to user settings and function control of the system.
  • the user input module 830 includes a touch panel and / or other input devices.
  • the output module 840 includes a display panel for displaying information input by the user, information provided to the user, various menu interfaces of the system, and the like.
  • the display panel may be configured in the form of a liquid crystal display (liquid crystal) (LCD) or an organic light-emitting diode (OLED).
  • the touch panel may cover the display panel to form a touch display screen.
  • the output module 840 may also include an audio output module, an alarm, and a haptic module.
  • the audio and video input module 860 is used to input audio signals or video signals.
  • the audio and video input module 860 may include a camera and a microphone.
  • the power supply 880 may receive external power and internal power under the control of the processor 850 and provide power required for the operation of various components of the system.
  • processor 850 may represent multiple processors.
  • the processor 850 may include a central processing unit (CPU) and a graphics processing unit (graphic processing) unit, GPU).
  • CPUs generally have multiple cores and are multi-core processors. The multiple cores can be integrated on the same chip or can be independent chips.
  • the memory 870 stores computer programs including an operating system program 872, application programs 871, and the like.
  • Typical operating system such as Microsoft's Apple's Developed by Google Android System etc.
  • the method provided by any one of the foregoing embodiments of the present application may be implemented by software, and may be considered as a specific implementation of the operating system program 872. As a part of the operating system program 872, it is deployed on a smartphone.
  • the processor 850 reads the operating system program
  • a function of the method operating system program 872 provided by the present application can be run in time.
  • the functions provided in this application can be enabled by default or can be enabled under the instruction of the user.
  • the operating system program 872 corresponds to the operating system 120 in FIG. 1; the application program 871 corresponds to the application 110 in FIG. 1.
  • the memory 870 may be one or more of the following types: flash memory, hard disk type memory, micro multimedia card type memory, card memory (such as SD or XD memory), random access memory (random access memory) , RAM) (such as the DDR SDRAM mentioned in the previous embodiment), read-only memory (read only memory (ROM), electrically erasable programmable read-only memory (electrically erasable programmable-read-only memory (EEPROM), programmable Read only memory (programmable ROM, PROM), rollback protected memory block (replay protected memory (RPMB), magnetic memory, magnetic disk or optical disk.
  • flash memory such as SD or XD memory
  • RAM random access memory (random access memory)
  • RAM random access memory (random access memory)
  • ROM read only memory
  • EEPROM electrically erasable programmable read-only memory
  • PROM programmable Read only memory
  • RPMB rollback protected memory block
  • magnetic memory magnetic disk or optical disk.
  • the storage 870 may also be a network storage device on the Internet, and the system may perform operations such as updating or reading the storage 870 on the Internet.
  • the processor 850 may indicate one or more processing units, for example, the processor 850 may include one or more central processors, or include a central processor and a graphics processor, or include an application processor and a coprocessor (Eg micro control unit or neural network processor). When the processor 850 includes multiple processing units, the multiple processing units may be integrated on the same chip, or may be independent chips.
  • a processor may include one or more physical cores.
  • the memory 870 also stores other data 873 in addition to computer programs.
  • the NPU 890 is mounted on the main processor 850 as a coprocessor, and is used to execute tasks assigned to it by the main processor 850.
  • the NPU 890 may implement some complex algorithms involved in the management and control decision process under the scheduling of the main processor 850.
  • the NPU 890 can also be used as the main processor to execute all the processes of the method provided by this application, or call its coprocessor to jointly implement the method provided by this application.
  • connection relationship of each module shown in the figure is only an example, and the method provided by any embodiment of this application can also be applied to smartphones with other connection methods, for example, all or some modules are connected by a bus To achieve communication.
  • FIG. 12 is a schematic structural diagram of an NPU 900, which is equivalent to the NPU 890 in FIG. As shown in the figure, the NPU 900 is connected to a main processor and an external memory, where the main processor is equivalent to the processor 850 in FIG. 11.
  • the core part of the NPU 900 is the arithmetic circuit 903, and the arithmetic circuit 903 is controlled by the controller 904 to extract data in the memory and perform mathematical operations.
  • the arithmetic circuit 903 internally includes multiple processing engines (process engines, PE). In some implementations, the arithmetic circuit 903 is a two-dimensional pulsating array. The arithmetic circuit 903 may also be a one-dimensional pulsating array or other electronic circuit capable of performing mathematical operations such as multiplication and addition. In other implementations, the arithmetic circuit 903 is a general-purpose matrix processor.
  • the arithmetic circuit 903 takes the data corresponding to the matrix B from the weight memory 902 and caches it on each PE of the arithmetic circuit 903.
  • the arithmetic circuit 903 takes matrix A data and matrix B from the input memory 901 to perform matrix operation, and the partial result or final result of the obtained matrix is stored in an accumulator 908.
  • the unified memory 906 is used to store input data and output data.
  • the weight data is directly transferred to the weight memory 902 through the storage unit access controller 905 (for example, direct memory access controller (DMAC)).
  • the input data is also transferred to the unified memory 906 through the storage unit access controller 905.
  • DMAC direct memory access controller
  • the bus interface unit 910 (bus interface unit, BIU) is used for the interaction between the AXI (advanced advanced interface) bus and the storage unit access controller 905 and the instruction fetch memory 909 (instruction fetch buffer).
  • the bus interface unit 910 is used to fetch the instruction memory 909 to obtain instructions from the external memory, and also used by the storage unit access controller 905 to obtain the original data of the input matrix A or the weight matrix B from the external memory.
  • the storage unit access controller 905 is mainly used to carry the input data in the external memory to the unified memory 906 or the weight data to the weight memory 902 or the input data data to the input memory 901.
  • the vector calculation unit 907 usually includes a plurality of operation processing units. If necessary, the output of the operation circuit 903 is further processed, such as vector multiplication, vector addition, exponential operation, logarithm operation, and / or size comparison, and so on.
  • the vector calculation unit 907 can store the processed vector into the unified memory 906.
  • the vector calculation unit 907 may apply a non-linear function to the output of the operation circuit 903, such as a vector of accumulated values, to generate an activation value.
  • the vector calculation unit 907 generates normalized values, merged values, or both.
  • the processed vector can be used as the activation input of the arithmetic circuit 903.
  • the fetch memory 909 connected to the controller 904 is used to store instructions used by the controller 904.
  • the unified memory 906, the input memory 901, the weight memory 902, and the fetch memory 909 are all on-chip memories.
  • the external memory in the figure is independent of the NPU hardware architecture.
  • the algorithm for searching the optimal scheduling strategy may be executed by the NPU 900, and then the execution result is returned to the main processor, thereby improving the method operating efficiency.
  • the NPU is only an example, and some algorithms in the solution provided by this application may be processed by other types of dedicated processors. Because the processing efficiency of the dedicated processor in processing specific algorithms is higher than that of ordinary processors, the use of some algorithms in the dedicated processor processing scheme can effectively improve the overall operating efficiency of the scheme, so as to find the appropriate resource management and control strategy as soon as possible and achieve The resource control strategy.
  • the method provided in this embodiment can also be applied to non-terminal computer equipment, such as a cloud server.
  • the device embodiments described above are only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, may be located in One place, or it can be distributed to multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement without paying creative labor.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Human Computer Interaction (AREA)
  • Telephone Function (AREA)

Abstract

本申请提供一种资源调度方法、装置及计算机设备等。该方法以提高用户的整体体验为目标确定并执行资源管控策略，其中，用户的整体体验由设备当前使用场景下性能、温度和功耗三者的权重以及三者各自的用户体验确定出来。采用这种融合性能、温度和功耗三种因素的调度方法，能够避免单一因素调度可能带来的跷跷板效应，提升用户对设备的综合体验。

Description

资源调度方法和计算机设备 技术领域
本申请涉及计算机技术，尤其涉及一种资源管理或者说资源调度的方法、装置、计算机设备、计算机存储介质和计算机程序产品等。
背景技术
以智能手机为代表的终端设备中，随着芯片性能的高速发展，终端性能大幅提升，然而电量(相当于待机时长)和发热成为降低用户体验的核心问题。如何有效地平衡性能、功耗和发热成为提升终端设备的竞争力的核心。
以安卓
Figure PCTCN2019104292-appb-000001
手机为例，当前是性能、功耗、热三者的管控独立进行。在性能管控方面，中央处理器(central processing unit,CPU)、图形处理器(graphic processing unit,GPU)、双倍速率(double data rate,DDR)存储器的频点仅基于历史负载进行调整。以CPU为例，当某个任务的CPU负载上升时，CPU频点上升。在功耗管控方面，针对不同的应用类型(如即时通讯类应用)对后台任务的执行时间和功耗消耗进行限制；或者根据整机剩余电量进行分级的功耗管控，例如剩余80％和20％电量对应不同的功耗管控级别。在热管控方面，根据终端设备上各个热区是否超过阈值做热源(如SOC、显示器背光灯)限制，以降低热源的频点、档位为代价限制温度上升。可见，性能、功耗、热是根据各自维度的劣化程度，独立进行调度，容易造成跷跷板效应，例如，电量充足时，优先调控性能，导致电量迅速降低，手机持续发热；电量不充足时，强制管控功耗和热，导致性能降低，影响用户体验。
发明内容
以下从多个方面介绍本申请，容易理解的是，该以下多个方面的实现方式和有益效果可互相参考。
第一方面，本申请提供一种资源调度方法，该方法包括确定设备当前的应用场景，其中，所述应用场景与用户的使用或设备的运行状态相关，运行场景通常包括以下多个维度中的任意一个或多个维度的信息：时间、空间、前台应用的类型以及应用内的使用场景；针对该应用场景，性能、功耗和温度分别有各自的权重，基于三者的权重以及三者目前的用户体验确定资源管控策略；根据所述资源管控策略调度所述设备的资源。
其中，性能、功耗和温度分别意味着三者各自的用户体验。用户体验有多种实现方法，比如对于功耗而言，用户体验的主要指标可以是待机时长，那么预估的待机时长越长，则功耗的用户体验越高；对于性能而言，用户体验的指标可以是视频的目标帧率、流畅度等，流畅度越高，用户体验越高；对于温度而言，用户体验的指标可以是手机外壳温度，手机外壳温度越高，则用户体验越低。性能、功耗和温度三者的权重分别指示三者对用户整体体验的影响程度。用户整体体验例如可以是性能、功耗和温度各自的用户体验的加权和。在确定资源管控策略的过程中，融合性能、功耗和温度三者各自的用户体验以形成用户的整体体验，以用户的整体体验为目标确定资源调度策略。
性能、功耗和温度，这三者很难融合，比如提升性能，功耗就会变高，同时温度就会上升，因此很难找到最合适的资源调度策略。而本申请提供的方法考虑到了不同的应用场景下三者对用户的重要性是不同的。针对特定的应用场景，融合性能、功耗和温度三者的权重确定合适的资源管控策略，然后根据该资源管控策略调度资源，从而实现了三种因素综合考虑的融合决策过程，有效避免了三者独立调度带来的跷跷板效应。同时，本申请提供的方法将性能、功耗和温度三个维度的用户体验满足程度的加权和作为用户整体体验，始终将用户整体体验的提升作为调度目标，有效避免了仅提升一个或两个维度的用户体验，而另一维度的体验下降过多，从而造成的用户整体体验降低的问题。
在一些实现方式中，资源管控策略的确定是周期性的，间隔时间段可以通过机器学习获得，也可以由设备的用户设置，或在设备出厂时预置在设备内部。
在一些实现方式中，资源管控策略可以由事件触发。具体的，当性能劣化程度、功耗劣化程度以及温度劣化程度中的任意一个或多个达到预设条件时，触发所述资源管控策略的确定。其中，劣化程度用于指示目标指标与实际指标之间的差距。通过事件触发资源管控策略的确定(或者说更新)，能够更加及时地发现设备的问题，并尽快提供新的资源管控策略。
在一些实现方式中，性能劣化程度用于指示所述设备的性能参数值与目标性能指标之间的差距，所述目标性能指标为满足所述应用场景下的性能体验的一个或多个性能参数的值；所述功耗劣化程度用于指示预测的可用电量与最低可用电量之间的差距，所述温度劣化程度用于指示预测的外壳温度和温度阈值之间的差距。
在一些实现方式中，所述目标性能指标由以下方式确定：在预先设置的数据库中搜索与所述应用场景匹配的目标性能指标，所述数据库中存储有多种应用场景以及多种目标性能指标的对应关系。该数据库可以存在本地，也可以存在其他设备上。例如本设备是手机，数据库可以存在手机上，也可以存在云端服务器上。在其他一些实现方式中，目标性能指标也可以通过在线估计的方式获得，或通过云端服务器计算再返回结果。
不同的应用场景对应不同的目标性能指标，通过在数据库中预先存储这种对应关系，需要时即可查询，从而快速地确定当前应用场景的目标性能指标，进而提升性能劣化程度的确定速度。
在一些实现方式中，设备的性能参数值采集得到的设备当前的性能参数值。在另一一些实现方式中，是预测得到的下一个时刻的性能参数值，具体的，根据所述应用场景和所述设备当前的系统状态可以预测得到设备的性能参数值。所述系统状态指示所述设备当前的资源使用情况，例如资源供给信息和系统负载信息。以上两种方式可以都执行，之后在获得的两组性能参数值中选择一个，或者取两者的平均值，或加权处理等，最终获得一组性能参数值与目标性能参数比较，以获得性能劣化程度。
在一些实现方式中，功耗劣化程度是由所述设备的电量使用规律以及所述设备的剩余电量确定的，其中，所述电量使用规律反映历史上的充电情况和耗电情况(或者说所述电量使用规律代表所述设备电量增加和消耗的历史)。所述电量使用规律包括充电周期、用电规律和应用使用规律。其中，所述充电周期代表历史上的充电间隔，例如充电开始时刻到下一次充电开始时刻、或充电结束时刻到下一次充电结束时刻、或充电结束时刻到下一次 充电开始时刻、或充电开始时刻到下一次充电结束时刻等。所述用电规律指示历史上多个时刻以及每个时刻的剩余电量。所述应用使用规律反映的是历史上应用对电量的消耗信息，可以包括多种信息，例如常用应用、单个应用的平均使用时长、单个应用的平均使用功耗和/或应用使用时所述终端设备所在的位置等反映应用的使用时长和使用功耗的信息。
换句话说，本申请提供了一种方法，利用该方法学习用户的电量使用规律，包括充电规律和应用对电量的消耗规律等，然后根据用户的电量使用规律来判定功耗可能的劣化程度，而非仅预置一些固定的待机时长阈值。这种方式使得功耗劣化程度的确定符合用户用电习惯，从而更加准确和有效。
在一些实现方式中，所述外壳温度是根据所述设备的当前外壳温度、所述设备内多个器件各自的温度升高预测值确定的。由于每个器件对设备外壳温度提升的贡献程度不一样，所以在计算外壳温度时通常还需要考虑每个器件温度的权重，该权重用于指示对外壳温度提升的贡献程度。
需要说明的是，这里的“器件”也可以理解为“热区”。一个热区可以包含一个或多个器件，温度衡量可以以单个器件为最小单元，也可以以热区为最小单元。热区是哪些，通常以设备内部预先设置的为准。
在一些实现方式中，温度劣化程度的计算中用到的温度阈值是根据所述设备当前所处的环境的温度确定的。环境温度的高低会影响用户的触感温度，所以用环境温度来决定温度阈值，更能贴合设备用户实际的温度体验，从而使得温度劣化程度的计算更加准确。
在一些实现方式中，所述资源管控策略包括以下三项中的任意一项或多项：计算资源调度策略、应用管控策略以及外设管控策略。资源管控策略是全方位的。
在一些实现方式中，所述资源管控策略的确定过程利用机器学习或动态规划的方法。充分利用机器学习等数学方法加快资源管控策略的确定过程。
在一些实现方式中，通常所说的设备的应用场景一般包括时间维度、空间维度、以及应用类型维度，比如14:00在家看视频应用。但是应用场景还可以更加细粒度。应用场景还可以包括应用内的使用场景。应用内的使用场景包括以下两项信息中的一项或两项：用户对应用的操作行为和应用内的进程行为。例如在微信
Figure PCTCN2019104292-appb-000002
上发朋友圈或看视频、在玩游戏的团战等。同样是玩游戏，可能单人作战和团队战场景对各个维度的需求是不一样的，细粒度的应用场景的划分能够使得用户体验目标的确定更加准确，从而使得确定出的资源管控策略更加有针对性，也更加有效。
在一些实现方式中，在确定资源管控策略的过程中，首先根据设备的性能瓶颈或主导热区确定候选的资源管控策略，然后再从候选的资源管控策略中确定待执行的资源管控策略。所述候选的资源管控策略能够提升所述性能瓶颈或降低所述主导热区的温度。通过这种筛选的方式可以减小资源管控策略的搜索范围，降低计算量，达到提升方法效率的目的。
在一些实现方式中，确定所述性能、功耗和温度各自的权重的因素包括以下任意一个或多个：所述设备的应用场景、所述性能、功耗和温度各自的劣化程度、以及所述性能、功耗和温度劣化判断的作用时域，作用时域例如可以是毫秒级、小时级或分钟级。
在一些实现方式，该方法中涉及的一些复杂计算，例如资源调度策略的确定过程，可通过专用处理器实现。专用处理器可作为协处理器，在主处理器的调度下执行。专用处理 器包括但不限于神经网络处理器、机器学习处理器等。另一方面，本方法中的任意步骤，即使不复杂，也可以通过专用处理器或微处理器或协处理器处理器。
第二方面，本申请提供一种资源调度方法，该方法包括：确定设备当前的应用场景；根据所述应用场景，融合性能、功耗和温度三者各自的用户体验以获得用户整体体验；根据所述用户整体体验确定并执行资源管控策略。
在一些实现方式中，根据性能、功耗和温度三者各自的用户体验以及三者在所述应用场景下分别的权重获得用户整体体验，例如三者各自的用户体验的加权和作为用户整体体验。其中，所述权重用于表示每个因素对用户整体体验的影响程度，在不同的场景下可能存在不同，例如游戏场景下用户更关注性能，所以性能的权重高于功耗和温度。
在一些实现方式中，根据性能劣化程度、功耗劣化程度、温度劣化程度分别确定所述性能的体验值、所述功耗的体验值和所述温度的体验值，其中，所述体验值用于指示用户体验，所述劣化程度用于指示目标指标与实际指标的差距，劣化程度越高，对应的体验值越低。然后根据所述性能、功耗和温度的三个体验值以及三者在所述应用场景下分别的权重确定用户整体体验。
第二方面的其他实现方式以及具体实现方式均可参考前述第一方面，在此不再赘述。
第三方面，本申请提供一种资源调度方法，该方法包括：基于用户的指令确定设备当前的应用场景；根据所述应用场景，融合性能、功耗和温度三者各自的用户体验以获得用户整体体验；根据所述用户整体体验确定并执行资源管控策略。
换句话说，本申请提供的方法可以在设备中处于默认关闭状态，由用户通过指令开启。
在一些实现方式中，设备的应用场景识别可以默认开启，但后续融合调度的方法可以由用户指令开启。进一步的，性能、功耗和温度三种模型可以分别设置开启指令，用户可以根据需求开启其中的任意一个或两个。
第三方面的其他实现方式以及具体实现方式均可参考前述第一方面或第二方面，在此不再赘述
第四方面，本申请提供一种计算机设备，例如终端设备，该设备包括处理器和存储器。所述存储器用于存储程序。所述处理器用于读取并执行所述程序并实现如前述任意一个方面或实现方式所描述的方法。
第五方面，本申请提供一种资源调度装置(或称之为资源管理和控制装置)，该装置包括一个或多个模块，该一个或多个模块用于实现前述任意一个方面或实现方式所描述的方法。
第六方面，本申请提供一种计算机程序(或计算机程序产品)，该计算机程序(或计算机程序产品)中存储程序，该程序再被一个或多个处理器读取并执行后实现如前述任意一个方面或实现方式所描述的方法。
第七方面，本申请提供一种计算机存储介质，可以是非易失性的。该计算机存储介质中包含程序，该程序再被一个或多个处理器读取并执行后实现如前述任意一个方面或实现方式所描述的方法。
附图说明
为了更清楚地说明本申请提供的技术方案，下面将对附图作简单地介绍。显而易见地，下面描述的附图仅仅是本申请的一些实施例。
图1为本申请提供的智能手机的层次划分示意图；
图2a为本申请提供的资源管理和控制装置的逻辑结构示意图；
图2b为本申请提供的资源调度方法的流程示意图；
图3为本申请提供的资源管理和控制装置的另一种逻辑结构示意图；
图4为本申请提供的性能模型的执行过程示意图；
图5为本申请提供的用户用电规律的一个示例；
图6为本申请提供的温度模型的执行过程示意图；
图7A和图7B为本申请提供的环境温度计算方法的流程示意图；
图8为本申请提供的资源调度方法的整体流程示意图；
图9为本申请提供的应用场景的维度示例；
图10为本申请提供的一种涉及策略筛选和反馈调整的流程示意图
图11为本申请提供的一种计算机设备的逻辑结构示意图；
图12为本申请提供的一种NPU的逻辑结构示意图。
具体实施方式
现有的性能管控机制，例如CPU的能量感知调度(energy-aware scheduling,EAS)和基于频点调节的governor机制，仅仅基于负载，存在性能供给过多，导致功耗和热控劣化的问题。以视频播放为例，视频播放的帧数不是固定的60帧，而是24或25或30帧。当前的性能管控机制对前台计算资源进行供给时没有考虑不同的帧数，存在资源供给过高导致功耗过高的问题。另一方面，CPU、GPU、DDR器件调度策略各行其是，互相影响，容易导致跷跷板效应。例如，电量充足时，优先调控性能，导致电量迅速降低，手机持续发热；电量不充足时，强制管控功耗和热，导致性能降低，影响用户体验。热控策略简单直接，发现温度升高就直接采用限制频率的策略。相同的计算量，频率被限制后导致计算时间被拉长，进一步引发了CPU功耗增加，并引发热量增加，热控效果被抵消。例如，游戏场景中，DDR调频策略使得DDR低频点，引发CPU负载升高，进而CPU频点上升，然而此时性能的瓶颈其实是内存，瓶颈导致的帧率过低的问题没有解决，反而由于CPU升频带来了发热问题。
有技术提出将功耗和热的管控进行联合，建立功耗和温升的对应关系模型，根据功耗预测温升。从温升预测出发，定期制定功耗预算，目的是满足智能手机温度低于温度阈值。将该功耗预算作用于热源，以达到降低功耗和控制温升的目的。该技术的本质上是通过功耗控制达到热控的目的，仅考虑了热和功耗的协同。但是这种方法很难保证用户的前台使用体验不会受到影响，依然存在为功耗或热控而降低前台性能的问题。进一步的，在前述功耗和热联合管控的基础上，有技术提出可让用户设置性能阈值，作为功耗和热的管控底线。即便如此，该技术对用户体验的考虑只是允许特定场景下用户手动标定不可接受的性 能底线，并没有把手机的整体使用体验考虑在内，例如待机时长、壳温等，容易导致手机整体使用体验降低。
以上技术提出的性能、功耗和温度的管控策略均没有针对用户使用模式差异针对性设置。不同用户使用手机的模式千差万别。例如，有些用户随时可以充电，待机时长不是体验瓶颈；有些用户长期出差在外，一天才能充一次电，对待机时长重视，不太看重游戏性能。再例如，低纬度地区环境温度高，用户对手机的发热很敏感，而高纬度地区，环境温度低，用户对发热耐受度很高。
因此，本申请提供提出了一种基于性能、功耗和温度(即发热或热)的融合资源调度方法。本申请的出发点是用户的使用场景(相当于终端设备的应用场景)，而不是性能、功耗和温度的独立负载(或劣化程度)，体现了不同使用场景下体验的侧重点差异。例如，游戏场景对性能的体验要求高，可以降低待机时长的体验限制。本申请的调度决策将性能、功耗和温度结合，把用户的综合体验最优化作为调度的目标，而不是从一个或两个目标出发产生决策点，避免了管控机制彼此的冲突。
进一步的，本申请提供的方法考虑了场景和用户的差异性，不是千人一面千篇一律地执行调度。具体的，本申请基于性能、功耗和温度三方面建立用户整体体验模型，针对不同的使用场景，该模型的侧重点有所差异；同时性能、功耗和温度三者的负载分别有各自的测量模型，这些测量模型是根据用户或设备的个体特点建立的。以上结合，最终生成基于不同场景的、且具备用户差异的调度策略。
下面将以智能手机为例介绍本申请提供的融合资源调度(或管理或控制)方法。应理解的是，本申请提供的方法不仅可以应用在智能手机上，还可以应在其他类型的终端设备上。这里的终端设备包括但不限于智能手机、车载装置、个人计算机、人工智能设备、平板电脑、个人数字助理、智能穿戴式设备(例如智能手表或手环、智能眼镜)、智能语音设备(例如智能音箱等)、以及网络接入设备(例如网关等)等。另外，本申请提供的方法还可以应用在服务器等其它类型的计算机设备上。
如图1所示，为本实施例提供的智能手机100的逻辑结构示意图。智能手机100包括三层：应用110、操作系统120以及硬件130。应用110包括各种应用，例如游戏、微信、YouTube视频、浏览器等。硬件130包括中央处理器(central processing unit,CPU)、应用处理器(application processor,AP)、图形处理器(graphic processing unit,GPU)、存储器(例如DDR)、输入设备、显示设备(例如触摸屏是输入与显示合一的设备)、各种类型的传感器等。有的未在图中示出。更多硬件可参考后续实施例的描述。
操作系统120(operating system，简称OS)是管理和控制计算机硬件与软件资源的计算机程序，是直接运行在“裸机”上的系统软件，应用110内包含的各种应用软件通常都需要在操作系统的支持下才能运行。操作系统120是用户和计算机的接口，同时也是硬件130和应用110的接口。操作系统120的功能包括管理硬件130内的各种硬件、应用110内的各种软件、以及各种数据资源，控制程序的运行，改善人机界面，为应用提供支持，让计算机系统所有资源最大限度地发挥作用。由此可见，资源管理和控制机制是操作系统120的核心(本实施例中将实现该机制的软件程序称之为资源管理和控制装置)。本实施例提出的资源调度方法正是运行在操作系统120内部的一种方法。通过该方法感知上层的应用场 景，监控底层资源运行状况，并融合性能、功耗、温度三个因素确定资源的调度策略，以提升智能手机100的整体体验。
图2a和图2b分别示出了本实施例提供的资源调度方法的一种模块划分示意图和方法流程的概要示意图。如图2a所示，本实施例提供的资源调度方法运行在操作系统120内部，主要包括三个模块：场景感知模块121、资源监控和决策模块122以及执行模块123。这三个模块分别执行图2b中示出的三个步骤S201-S203。
S201、确定手机当前的应用场景。
应用场景通常包括以下一个或多个维度的信息：时间、空间、前台应用的类型以及应用内的使用场景。例如，“在家看视频”这个场景包括空间和前台应用的类型两个维度。
手机的应用场景可以理解为用户使用手机的使用场景，比如开会场景、睡眠场景、工作场景、家场景、噪音场景、开车场景、跑步场景等。
手机的应用场景可以是单维度的，也可以是多维度的，例如用户在噪音环境下看视频。进一步的，若用户正在使用某个应用，可以在该应用内继续细分不同的场景，例如正在观看1080p或720p的视频，或者正在玩某个游戏中的团队战役等。
需要说明的是，手机的动作或运动状态可以划归到“空间”维度。举例说明，跑步场景的前台应用类型可能为某个健康应用或某个音乐应用，同时手机上的感应装置感应到手机在规律性运动，即“跑步场景”包含前台应用的类型和空间这两个维度的信息。
S202、基于所述应用场景下性能、功耗和温度三者的权重确定资源管控策略。
确定资源管控策略的触发条件可以是性能劣化程度、功耗劣化程度以及温度劣化程度中的任意一个或多个达到预设条件。
性能劣化程度用于指示所述设备的性能参数值与目标性能指标之间的差距，所述目标性能指标为满足所述应用场景下的性能体验的一个或多个性能参数的值。功耗劣化程度用于指示预测的可用电量与最低可用电量之间的差距。温度劣化程度用于指示预测的外壳温度和温度阈值之间的差距。
需要说明的是，本实施例中所说的两项的“差距”指该两项的区别。“差距”的实现方式有多种，例如可以是这两项的差值、这两项的比值，或在差值或比值的基础上再增加其他的因素后获得的值；或者不是具体的值，是一种表现趋势的函数，例如性能参数值的发展与目标性能指标相违背，比如性能将要下降，但目标性能指标为更高值。换句话说，只要能够表征出劣化程度即可，本实施对具体实现方式不做限定。
这里的性能、功耗和温度可以分别理解为性能体验、待机时长(即功耗体验)、外壳温度(即温度体验)。根据以上三者分别的权重确定资源管控策略。该权重和前一步骤确定的手机应用场景相关，不同应用场景下，三者的权重通常不同，例如游戏场景下性能体验的权重高于温度和功耗。
具体的，首先根据前一步骤确定的应用场景确定目标性能指标。然后计算所述目标性能指标与手机当前配置的性能参数值之间的差距以确定性能劣化程度，并根据所述性能劣化程度确定是否生成性能劣化告警。根据所述应用场景、电量使用习惯以及所述终端设备的当前剩余电量确定功耗劣化程度，并根据所述功耗劣化程度确定是否生成功耗劣化告警。根据采集到的热区温度预测所述终端设备的外壳温度，根据所述外壳温度确定温度劣化程 度，并根据所述温度劣化程度确定是否生成温度升高告警。当所述性能劣化告警、功耗劣化告警以及所述温度升高告警中的任意一个或多个被生成时，说明当前的资源管控出不能满足手机整体的用户体验，因此开始资源管控策略的确定。
在其他实施例中，用户可以设置资源管控周期，周期性触发策略确定过程。
S203、根据所述资源管控策略调度所述设备的资源。所述资源管控策略包括以下三项中的任意一项或多项：计算资源调度策略、应用管控策略以及外设管控策略。
在一些实施例中，前一步骤确定的资源管控策略为满足所述应用场景下用户体验最优的资源管控策略。用户体验由性能、待机时长和温度三者的满足程度以及三者分别在所述应用场景下的权重通过一定的数学变换获得。
确定资源管控策略的过程可以理解为求解用户体验最优的数学过程，例如机器学习或动态规划的方法。在其它一些实施例中，“最优”未必理解为理论上或数学上的最大值。用户体验作为目标体验，该值可以由用户设置，或者该值虽不是数学上的最大值，但可能是当前场景下最合适的一个值。
由以上可见，通过本实施例提供的方法，手机可以感知用户所处的使用场景，根据使用场景确定目标资源需求(相当于目标性能指标)，并计算性能需求和当前资源配置之间的差距，以获得性能劣化数据，同时结合功耗和热这两方面产生的劣化数据，基于性能、功耗和热这三方面在当前使用场景下的权重产生调度决策并下发，以达到同时保证前台性能、待机时长和手机壳温处于稳态的目的。
与现有技术相比，本申请提供的方法在识别用户使用场景的前提下，将性能、功耗和热三者按照当前场景下的权重结合起来，以三者结合的用户综合体验最优化作为调度的目标，而不是从一个或两个目标出发产生决策点，避免了管控机制彼此的冲突，体现了不同使用场景下三者的侧重点，从而生成更加有效的资源调控策略。
图3为本实施例提供的资源管理和控制装置进一步的模块划分示意图，尤其是对于资源监控和决策模块进行了更加详细的描述。如图所示，该装置包括场景感知模块210(相当于场景感知模块121)、执行模块260(相当于执行模块123)、以及220-250四个模块。
场景感知模块210用于识别终端的应用场景。该应用场景由一个或多个维度的信息标识。例如用户使用情景(家/办公室、白天/黑夜、工作日/休息日)、应用类型(如视频类、游戏类等)、应用内细粒度使用场景(如在看1080p高清视频)等。
性能模型220用于预测性能是否会劣化或性能劣化程度。具体的，性能模型220用于感知不同应用场景下的目标性能目标，结合当前的系统状态(包括资源供给状况和系统负载情况)，预测和判断用户性能维度的体验是否会劣化以及劣化的程度。进一步的，性能模型220会根据芯片上报的PMU(performance monitor unit)信息，分析和识别出造成性能劣化的主要瓶颈。
需要说明的是，性能模型的预测通常是毫秒级别。
功耗模型230用于预测功耗劣化程度。在一些实施例中，根据当前应用、手机剩余电量和学习出的用户用电习惯，预测未来的一段时间内(例如1小时)是否满足待机时长的约束，若不满足约束则给出功耗劣化的程度。功耗劣化程度可以用待机时长或者电量劣化程度来表示。
需要说明的是，功耗模型的预测通常是小时级别。
热控模型240用于预测温度劣化程度。在一些实施例中，根据热稳态算法和壳温预测算法，判断未来的一段时间(例如8分钟)是否会出现手机温度超过阈值，引起严重发热的现象，若存在会发出只是。进一步的，热模型会根据手机上的热区的温度和热区对最终外壳温度的影响程度，来识别导致发热的主导热区。
需哟啊说明的是，热控模型的预测通常是分钟级别。
决策模块250用于基于当前应用场景下性能、功耗和温度三者的权重确定资源管控策略。在一些实施例中，根据来自于性能模型、功耗模型和热模型的告警信息触发融合调度决策，综合考虑上述三个维度的需求、现状、以及当前应用场景下三个维度的权重来动态决定用户综合体验最佳的最终方案。
执行模块260用于根据上述确定的资源管控策略实施资源管控。资源管控通常涉及多种资源的多项管控措施，因此调度管控模块可以理解为多个分布在系统的不同地方的功能单元的组合。资源管控包括对实时资源、后台进程和外设等的调度。
下面分别从场景感知、性能模型、功耗模型、热控模型、融合调度这几个角度详述本申请的多种实现方式。下面的各个方法可以理解为前述相应的功能模块的具体实现。容易理解的是，一个功能模块可以包含多个子模块，功能模块的具体实现可以是这多个子模块协作完成的。
(1)场景感知(对应场景感知模块210)
场景感知是多种识别方法的综合，输出用户的具体使用场景，如当前用户在家里在看清晰度1080p的硬解码视频。场景感知包括用户时空感知、应用类型感知、以及应用内场景识别三个方面。
用户时空感知又可以包括时间维度和空间维度。时间维度例如节假日/工作日、白天/黑夜。通过获取当前的系统时间或从系统的日历中查询以区分时间。空间维度例如出差/旅游、家/办公室。通过采集得到的GPS位置信息进行聚类，按照不同时间的最频繁出现的地理位置点识别家和办公室位置；结合当前的时间来确定是否出差或旅游等。
应用类型感知即感知用户在前台使用的是什么应用。可以在手机系统中预置多种应用类型的列表，例如国内使用人数排名前10000名的应用类型，然后根据当前的前台应用查询列表以确定该前台应用对应的应用类型，比如微信(WeChat)对应即时通讯类应用。在另一些实现方式中，手机也可以将当前应用的标识通过网络发送给云端服务器，由云端服务器识别并返回识别结果给手机。
应用内场景识别即识别应用内更细粒度的场景。应用内场景维度包括用户对应用的操作行为(简称用户行为)和应用内的进程的行为(简称进程行为)。用户的行为例如播放不同清晰度的视频、进行视频聊天等属于用户的行为，游戏的启动、放大招、选英雄等属于应用内进程的行为。用户行为可通过打点上报各种事件以及在数据流中执行元数据识别等方式获取。进程行为可以通过应用的绘制信息或预先嵌入在应用中的SDK主动向系统上报等方式识别。
(2)性能模型(对应性能模型220)
性能模型主要用于识别不同场景下的性能目标，并结合当前的系统状态(包括资源供给状况和系统负载情况)，预测和判断用户性能维度的体验是否会劣化(毫秒级别)，以及劣化的程度。进一步的，性能模型会根据当前使用场景和系统状态，分析和识别出造成性能劣化的主要瓶颈。
图3为场景识别以及确定性能劣化告警的方法流程示意图。该方法大致分为以下a)、b)和c)三部分。
a)识别场景和对应的目标性能指标
通过各种方式实现场景识别的过程，前述(1)中已经介绍了，在此不再赘述，可参考图4。识别场景(S310)之后，在预先设置的数据库301中搜索与所述场景匹配的目标性能指标(S320)。数据库301存储有多种应用场景以及多种目标性能指标的对应关系。该数据库可以是离线生成并预先配置的。
具体的，该数据库中包括不同场景(记做Si,i＝0…n)，以及不同场景对应的目标性能指标Ti。Ti包括：目标帧率、丢帧率、低帧率占比(低于目标帧率的帧数占总帧数的百分比)、以及流畅度(帧率波动情况)等。
在一些实施例中，数据库中的场景划分和体验目标可以是人工参与制定的。在另一些实施例中，数据库中的场景划分和目标性能指标是从通过对上报的大数据进行聚类发现的，例如，视频播放场景细分为不同的帧率(24/25/30帧)和软硬解码场景，分别对应不同的体验目标。
在系统运行时，场景感知模块210可以用于识别场景S i，然后性能模型220查询数据库301以获得该场景S i的目标性能指标T i。例如，场景感知模块210通过视频流数据的解码方式和帧头部信息确定是硬解码30帧场景，该场景输入性能模型220，性能模型220确定对应的体验目标可能为Ti＝{平均30帧，丢帧率低于30％，低帧率占比不高于20％，流畅度在0.8以内}。
b)确定系统当前状态下的性能指标
根据当前的应用场景和系统状态，预测下个采样周期(毫秒或秒级)的帧率变化并通过计算获得其他维度的性能指标(S330)。其中，系统状态包括资源供给状况和系统负载情况。其他维度的性能指标，例如丢帧率、低帧率占比以及流畅度，可基于当前帧率变化和相应的计算公式计算获得。计算公式如下：
Figure PCTCN2019104292-appb-000003
其中，
Figure PCTCN2019104292-appb-000004
是下个采样周期的帧率，
Figure PCTCN2019104292-appb-000005
是当前采样周期采集到的帧率，
Figure PCTCN2019104292-appb-000006
是预测的帧率变化，其具体计算方法为：
Figure PCTCN2019104292-appb-000007
其中，r表示资源的某个维度，R表示所有资源的集合，r cur表示当前资源供给状况， r within_scene表示在这个应用内场景下资源的需求情况，
Figure PCTCN2019104292-appb-000008
表示进行归一化操作，防止维度之间的值域相差过大引起的维度消失情况，α app表示不同应用在不同资源上的权重系数，β overload表示系统负载对应的权重系数。
其中α app和β overload可以通过如下方式获得：第一种，离线分析后预置在手机中；第二种，在线分析，例如统计应用在运行时的各个资源的使用情况，分析这个应用对不同资源的侧重/需求，基于此来学习各个维度的权重。
除了上述预测的方式可以获得系统当前状态下的性能指标之外，还可以通过数据采集的方式获得。参考图3，采集一段时间的帧率，并获得，并通过采集或计算的方式获得平均帧率、丢帧率、低帧率占比、流畅度等其他维度的性能指标(S340)。
需要说明的是，预测和采集的数据可以择其一使用，也可以对这两种方式获得的数据做一些处理后使用，例如取二者的平均值。
c)用户性能体验差距的评估
根据目标性能指标和当前实际反馈的性能指标，可以计算当前性能距离目标性能的差距，即性能劣化程度(S350)。这里，实际反馈的性能指标可以通过上面提到的预测方式所获取的，也可以通过实时采集的方式所获取的。
性能劣化程度的计算公式如下：
Figure PCTCN2019104292-appb-000009
其中，α i为第i项指标在不同场景下的劣化权重。∑α i＝1。例如，游戏场景中平均帧率劣化对应的权重就比视频场景下平均帧率的权重大。T target i表示第i项指标的目标值，如游戏中对战场景的平均帧率的目标性能指标是57帧。T out i表示第i项指标的实际值。
当ΔT超过阈值时(如超出一定的百分比)，视为性能发生劣化，需要将性能劣化告警事件上报给决策模块250。
需要说明的是，不同的性能指标有不同的要求，有的量化后数值越高越好，有的量化后的数值越低越好，所以公式
Figure PCTCN2019104292-appb-000010
是可以根据具体情况变化的，此属于常见的数学方法，本实施例在此不一一列举，具体示例可参考后续描述。
需要说明的是，以上步骤除场景识别由场景感知模块210执行之外，其他的步骤均可由性能模型220执行。
(3)功耗模型(对应功耗模型230)
功耗模型用于学习用户的电量使用习惯，并根据当前应用场景和当前的应用对待机时长或者可用电量的影响程度进行评估。对于确定的用户，待机时长和可用电量两个变量之间可相互转换，故下文将仅以可用电量作为例子来说明。功耗模型的具体实现大致包括以下a)、b)和c)三个部分。
a)采集和记录用户的电量使用信息
本步骤会在学习用户使用习惯之前执行，采集和记录的信息包括但不限于用户使用的应用、使用时长、充电开始时刻和充电结束时刻、充电地点、应用平均功耗等信息，并将其进行持久化存储，其存储形式为文件或数据库。
b)分析和学习用户的电量使用规律
本步骤以步骤a)中持久化保存的电量使用信息作为输入，通过机器学习或者统计学的方法来学习用户的电量使用规律，其中电量使用规律包括充电周期、用电规律和应用使用规律。
具体的，以充电开始时刻和充电结束时刻作为输入，学习出充电周期。不同用户充电周期可能不同，例如有的人一天一充，有的人两天一充。统计用户在每个时间点的剩余电量，学习出基于时间点的历史剩余电量，作为用电规律。以用户使用的应用、使用时长、应用平均功耗、用户位置等作为输入，学习出在不同位置和时间段的用户常用应用，以及应用的平均使用时长和平均使用功耗等信息作为应用使用规律。
c)基于使用习惯的功耗划分
以步骤b)中学习出的充电周期和用电规律作为输入，通过时间段划分和统计学方法，输出每个时间段及这个时间段对应的最低可用电量。
时间段以充电周期作为划分基础，而不是固定的一天。不同的用户充电周期有较大差异，例如有的人一天一充，即充电周期时长为24小时；有的人两天一充，即充电周期时长为48小时。
在一些实现方式中，时间段的划分方法可以是固定时间片，例如每3小时作为一个时间段。在另一些实现方式中，时间段的划分可以基于语义，例如早晨、上午、中午、下午、傍晚、晚上、半夜。在另一些实现方式中，时间段可以使用动态划分方法。
时间段的动态划分方法举例如下：先使用预置的时间段的下限个数，如3个，对充电周期进行等距划分。确定每个时间段的持续时长，若时长超过阈值，例如3小时，则在这个时间段进行插值，保证每个时间段的持续时长不超过时长阈值。评估每个时间段的电量变化或待机时长变化，若一个时间段的电量变化或待机时长变化超过阈值，例如上一个时间段最低可用电量为80％，后一个时间段最低可用电量为20％，电量变化为60％，超过阈值20％，则在这个时间段进行插值，保证每个时间段的电量变化不超过电量阈值。
图5为划分好的时间段的示意图，其中每个时间段的最低可用电量计算方法如下：
Figure PCTCN2019104292-appb-000011
其中，
Figure PCTCN2019104292-appb-000012
表示第i个时间段的最低可用电量，N表示过去发生的次数，
Figure PCTCN2019104292-appb-000013
表示第j个历史上该时间段的剩余电量，γ(·)表示弹性函数，即允许最低可用电量可以在γ(std,location,scene)所代表的值区间内进行波动，它以历史统计的标准差std、当前位置location和当应用场景scene作为变量。
d)根据当前应用场景的实时功耗劣化评估。
本步骤在步骤c)完成后执行。本步骤是周期性执行或者用户使用的应用发生变化时被触发执行。本步骤的输入是当前电量、步骤b)中学习出的应用使用规律和步骤c)计算得到的时间段划分信息，输出是功耗劣化评估和劣化程度。
具体的，根据应用使用规律和当前的前台应用，预测出该前台应用的使用时长以及下一个时间段所预估计的电量消耗ΔE，然后根据下一个时间段需要满足的最低可用电量
Figure PCTCN2019104292-appb-000014
评估是否出现功耗劣化。如果当前剩余电量E cur减去预估计的电量消耗ΔE低于下一个时间段需要满足的最低可用电量
Figure PCTCN2019104292-appb-000015
则评估会出现功耗劣化，并将告警发送给融合调度模块进行决策。劣化程度ξ的评估可通过下面的公示计算所得：
Figure PCTCN2019104292-appb-000016
(4)热控模型(对应热控模型240)
根据热稳态算法和壳温预测算法，判断将来的几分钟内是否会出现温度升高预警(手机外壳温度超过阈值)。同时，热控模型会根据热区温度和热区对最终温度的影响程度，来识别导致发热的主导热区。参考图6，热控模型的实现主要分为以下a)-e)几个部分。
a)散热能力识别
手机的布局，包括手机的结构布局和手机的外壳材质，会影响手机本身的散热能力。借助函数逼近方法，例如最小二乘方法，离线学习出多种使用场景下的该手机的散热能力曲线，并在设备出厂时提前进行预配置。
b)壳温预测
壳温就是手机的外壳温度。所述外壳温度是根据手机当前的外壳温度、手机内多个器件各自的温度升高预测值、以及每个器件温度的权重确定的。器件的温度升高预测值根据该器件的温升函数和该器件的当前资源配置确定，其中，所述温升函数可以预置在手机中。
具体的，根据采集的当前的热区温度，以及热区本身对壳温的影响程度来预测未来短期内t时刻(分钟级预测)的手机壳温，具体计算公式为：
H(t)＝T now+∑α i*F i(R i,t)    (6)
其中α i是各热区温度的权重，与硬件设备的布局等相关。T now是当前时刻对热区采样的温度值。各热区在未来T时刻的温升温度可用F i(R i,t)来度量。其中R i是当前的系统资源配置(或供给)情况，t是持续时间。F i是第i个热区器件(如CPU或GPU或DDR)的温升函数，其与产品的硬件配置和布局有关，一般是多个因素线性拟合的结果，为固定函数。
DDR即双倍速率，DDR SDRAM即双倍速率同步动态随机存储器，本领域习惯称为DDR，其中，SDRAM是synchronous dynamic random access memory的缩写，即同步动态随机存取存储器。
需要说明的是，一个“热区”可以仅包含一个器件，也可以包含多个相同类型或不同类型的器件。
c)主导热区识别
计算每个热区对壳温上升所贡献的比例并进行排序，选取排序在前的N个热区作为主导热区，贡献比例的计算方法为：
Figure PCTCN2019104292-appb-000017
其中ρ j是热区j对壳温所贡献的比例，α j*F j(R j,t)是热区j产生的壳温上升量，∑α i*F i(R i,t)是所有热区产生的壳温上升量。
识别出主导热区，可以为决策模块250的调度决策提供输入。
d)环境温度识别
环境温度(简称环温)计算的输入包含：NTC节点采集的温度数据、各器件的功耗数据以及温箱环境温度数据进行模拟训练，计算得到环境温度模型的拟合参数。流程如图7A所示。首先将当前系统状态及各个NTC节点采集的存入状态历史表(S701)，然后判断是否需要更新环境温度(S702和S703).如果与上次环温输出的间隔时间没有超过间隔阈值(S702)，则不更新，反之则继续判断是否处于充电状态(S703)，如果处于充电状态则不更新。如果与上次环温输出的间隔时间已经超过间隔阈值且不处于充电状态，那么对状态历史表中的温度数据进行过滤(1)，然后根据过滤之后的数据计算最新的环境温度(2)，计算结果可能需要一些过滤处理(3)，过滤处理后将最终的结果加入环温历史表(S704)。
NTC全称是thermistor temperature sensor。NTC节点就是指温度采集传感器，用于采集各个热区的温度。
环境温度的计算有一定的周期，当距离上次计算时间超过阈值，并且当前手机不处于充电状态时，执行最新环境温度计算，并将计算结果加入环境温度历史表。
如图7B所示，环境温度计算采用类Kalman滤波的计算方法，具体的，可以采用线性时不变降阶模型(linear and time-invariant reduced-order model,LTI-ROM)来预测环境温度，该模型的输入是多个功耗P1-P5以及采集的温度值。LTI-ROM模型输出的是下一个时刻的环境温度预测值，该预测值被输入校正模型，由校正模型结合环境温度的实际值对预测值进行校正，预测值和实际值一致的时候输出环境温度，不一致时反向调节LTI-ROM模型的参数。
e)温升告警决策
热控模型中，温升告警的判断条件可以包括以下两个：短期内的最大温度不能超过阈值，同时这段时间内温度波动范围不能超过阈值。
用公式描述：
Figure PCTCN2019104292-appb-000018
其中时间段T的整机温度用H(T)描述，T thresold是温度上限阈值，ε是温度波动阈值。上述公式的意义是时间段T内的最高温度小于T thresold，同时最高温度和最低温度的差值小于ε。在其他实施例中，“小于”也可以是“小于或等于”。
温度阈值T thresold和/或ε是可以根据环境温度调整的。例如环境温度很低，那么T thresold可以设置的高一些，因为此时人对手机外壳温度的容忍程度比较高；相反，环境温度很高，那么T thresold可以设置的低一些，因为此时人对手机外壳温度的容忍程度比较低。
在其他实施例中，上述两个条件并非缺一不可，可以选择其中任意一个作为温升告警的判断条件。
5)融合调度模型
当有性能劣化告警、功耗劣化告警或者温升告警中的一个或者多个事件出现时，会触发融合调度决策，并根据融合调度模型决定用户整体体验最佳的最终方案，对实时资源、后台应用和外设等进行管控和调度。
在当前状态X＝{手机使用场景*性能劣化程度*功耗劣化程度(相当于待机时长劣化程度)*温度劣化程度}的前提下，在可用的管控策略空间A中搜索最优的策略a，使得短期内用户的整体体验R最优。其中策略a的组成可以为{CPU/GPU/DDR档位,外设管控档位,应用管控策略}。用户的整体体验R是以下因素的加权值：
R＝∑α i*S i    (9)
其中S包括以下维度：性能体验目标满足程度(S 1)、待机时长(S 2)以及壳温体验(S 3)。换句话说，整体体验R与性能、功耗以及外壳温度三个因素相关。
α i是每个因素的权重，该权重在不同的终端应用场景下(或不同的用户)有不同。例如，游戏场景中性能体验目标满足程度的权重就要高于阅读电子书场景。搜索的目标就是寻找最大化R的策略a，即max a∈AR。
在具体实施中，可将融合调度决策的选择过程抽象化为一个强化学习或马尔科夫决策过程。以强化学习为例，
· 离线状态下，根据走访、用户体验，确定各种场景下体验目标权重α的选择；
· 将融合调度决策选择过程抽象为四元组E＝<X,A,P,R>。其中环境X表示当前的状态，即{使用场景、性能/功耗/热劣化程度}。A表示当前可选择的策略，如资源供给调整、外设管控策略、应用管控策略构成的策略组合。P是离线环境下学习得到的状态转移概率矩阵。R是在X状态下作出A策略的奖赏，即用户整体体验的优化/劣化变动；
· 通过离线场景学习的P和R，在运行时根据状态X，采取相应的策略A。
同时，融合调度决策会根据运行时的系统和用户的反馈来学习和更新后续的策略选择。涉及的反馈包括：
· 性能维度:丢帧、卡顿、外设主动调节反馈(调节背光、音量)、后续使用的应用等；
· 功耗维度：电池耗尽的次数、某个时间段功耗是否超标等；
· 温度维度：电池发热次数、电池温度是否有效下降等。
下面结合游戏场景的一个实例以及前述方法本申请提供的融合调度方法。
如图8所示，该方法主要分为场景感知、性能模型、功耗模型、热控模型四个感知模块，融合调度中心(含决策库)一个决策模块和一个对接外部调度机制的执行模块。以下是各个流程的具体描述。场景感知负责识别细粒度的使用场景。性能模型是毫秒级的感知模块，负责根据使用场景确定目标性能需求，以及确定当前性能指标，然后根据两者确定性能劣化程度和性能瓶颈。功耗模型是小时级别的感知模块，负责判断功耗(即待机时长)劣化程度并适时输出预警。热控模型是分钟级别的感知模块，用于判断温度升高程度以及主导热区。融合调度中心接收到性能模型、功耗模型、热控模型这三个感知模块的输入，然后做出整体性能最优的调度策略，并下发到执行模块。调度策略可以包括计算资源调度、后台应用管控以及外设管控等。决策库用于存储一些预置的调度策略，以供融合调度中心在其中搜索以获得整体性能最优的调度策略。
调度策略的执行过程并非本申请的重点，所以除执行模块之外，下面对剩余的五个流程分别进行详细的描述。
流程一.用户使用场景感知
对用户的使用场景(即手机的应用场景)进行感知，作为性能模型的基础输入，用于识别用户前台的性能需求。
如图9所示，用户的使用场景包括几个维度，分别由以下数据来源提供。
· 用户的时间维度和空间维度，包括当前时间、当前用户所在位置类型。时间维度可通过系统日历进行查询，例如节假日和工作日。空间维度包括家/办公室/其他，可通过通过位置聚合服务获得，例如某些手机提供了情景智能服务，则可以通过查询情景智能服务接口获取。
· 应用类型维度，如即时通讯类应用、邮件类应用或游戏类应用。应用类型一般表示应用的核心用途。例如微信(WeChat)的类型就是即时通讯类应用。数据的来源可以是查询应用类型数据库(数据库设置在端侧或云侧)，或在端侧动态识别，亦或从云端查询。
· 应用内场景维度，如微信内可以在朋友圈里看视频，则应用内场景就是视频播放场景。场景可根据框架(Framework)中打点上报的应用行为事件和前台应用关联信息进行识别。例如，根据打点上报的全屏事件以及视频播放事件(或视频类型应用和发声事件)可以识别视频播放场景。对于非游戏，还可通过应用活动(Activity)和场景关联，通过前台的Activity识别不同的应用内场景。例如，ActivityManagerService 上报com.tencent.mm/com.tencent.mm.plugin.sns.ui.SnsTimeLineUI这个Activity在前台，则可知用户当前正在使用微信应用的朋友圈功能。
· 应用内场景参数维度，如优酷(Youku)内看视频区分全屏/非全屏、软解视频/硬解视频、360p、720p/1080p等不同的清晰度、是否有用户字幕显示等。这部分数据可通过在框架中打点获取，或集成了特定sdk的应用主动发送相关的参数。
用户的使用场景感知结果，封装成scene数据类型表示的细粒度使用场景，传递给性能模型、功耗模型，作为这两种模型的输入。
以游戏为例，假设用户当前在玩G游戏，游戏的帧率有30、40、60帧的不同场景，则用户场景感知输出的scene信息为：<节假日，晚上，家里，G游戏，团战，60帧>。其中“G”指游戏应用的名称，“团战”指的是该游戏中的一种多人战争场景。
流程二.性能模型
性能模型用于发现性能体验的劣化程度，发出性能劣化感知信息和性能瓶颈，通知融合调度中心。对于游戏场景，性能体验指标通常包括以下维度：平均帧率、低帧率占比以及抖动率信息。其他场景下，性能体验指标可能与游戏场景相同，也可能不同，本实施例对此不做限定。
在性能模型中按照以下顺序执行1-4的顺序执行，最后输出性能劣化程度和性能瓶颈。
1.识别场景体验目标
根据Scene＝<节假日，晚上，家里，G游戏，团战，60帧>，在数据库中进行查询，获得当前场景对应的性能指标。
具体的，性能指标可以根据手机的分级进行区分。例如旗舰机、中端机、低端机分别对应性能目标A、B和C。三种手机级别在数据库中可以实现为三张表，每张表包含该手机级别的场景和性能指标的对应关系；也可以是一张表，由手机级别字段进行区分。
延续前述例子，从框架层上报信息得知运行环境为旗舰机，分辨率为1080p，则确定当前Scene对应的性能目标A就是<平均帧率57，低帧率占比0.5％，抖动率0.2％>。
如果是中端机，则对应的性能目标B为<平均帧率56，低帧率占比1％，抖动率3％>。如果是低端机，则对应的性能目标C为<平均帧率55，低帧率占比1％，抖动率1％>
2.采集前台应用性能
通过在Surfaceflinger中进行打点，采集应用当前的实际性能指标。除了当前的实际性能指标，性能模型根据系统当前的负载和SOC资源供给状况，可进行短期内性能指标变化预测。Surfaceflinger是安卓系统框架层的一个组件。SOC全称是system on a chip，可认为是芯片组，集成处理器的各个核，本实施例中包括4个大核和4个小核。在本实施中，GPU、DDR也在SOC上。
根据采集或预测得到的实际性能指标与性能目标的差距，可以发现性能劣化。对这种性能劣化进行量化，产生不同级别的性能劣化告警。
3.性能瓶颈发现
性能瓶颈发现主要依靠SOC上的性能监控单元(performance monitor unit，PMU)上报事件进行识别，为已有成熟方案，此处不进行赘述。
4.量化场景劣化
性能的劣化量化方式为对三项体验指标的加权，其量化方式可参见前述公式(3)。
延续前述例子，在G游戏团战场景中，三种指标的劣化程度权重为：平均帧率0.5、低帧率0.3、抖动率0.2，对应公式(3)中的α i。换句话说，平均帧率在G游戏团战场景中的权重最大；低帧率次之；抖动率的权重最小。
每个维度的劣化程度的量化定义为当前性能指标T out i距离性能目标T target i的百分比差距。例如性能目标A中的平均帧率目标为57帧，当前平均帧率为50帧，则平均帧率劣化值为(57-50)/57＝0.12。
例如，低帧率劣化值为(0.3-0.5)/0.5＝-0.4。其中，目标值为0.5，当前值为0.3，当前值优于目标值，所以劣化值是负数，此时可以强制设置劣化值为0。类似的方法，抖动率劣化值为(0.3-0.2)/0.2＝0.5，其中目标值是0.2，当前值是0.3，劣化值是0.5。
根据公式(3)ΔT＝0.5*0.12+0.3*0+0.2*0.5＝0.16。不同的ΔT值对应不同级别的性能劣化告警，将性能劣化告警上报到融合调度中心。具体的，可以直接上报0.16，也可以将0.16转化成性能劣化告警的级别，例如对应的级别为0.3，将0.3上报到融合调度中心。
流程二.性能模型
功耗模型在本实施例也需要考虑当前的使用场景，因为不同使用场景下的功耗可能是不一样的，比如本实施中的游戏场景，用电量的消耗就要快于聊天场景。
功耗模型的运作过程主要包括以下1和2两步。
1.学习用户的充电习惯和用电习惯，判断剩余待机时间。
功耗模型在后台收集电量使用信息，电量使用信息分为两部分：充电习惯和用电习惯。
a)充电习惯
统计当前时间和空间类型(工作日/休息日，家/办公室/其他，来自于场景识别模块)中，连续两次从80％以下充电到100％的时间之差，作为用户的充电记录。
对过去n次同类时空类型的充电记录进行统计学习，例如取n＝7即最近7次充电记录作为窗口，计算窗口内充电开始时间段的方差。例如，过去7次充电时间分别为9点、10点、9点、9点30、8点30、9点、10点半，则对应的方差为0.41。
如果充电记录的方差小于一定阈值，则认为用户的充电习惯稳定，可得到用户在当前时空类型下距离下一次充电的时间预测。如果充电记录方差比较大，则使用默认充电时长 (12小时)作为预测，计算距离下一次充电的时间间隔。
b)用电习惯
将一天划分为k个时间片，如取k＝8，每个时间片为3小时。统计在当前时间片内用户常用的应用、应用单位时间内耗电量，并加权得到该时间片内用户功耗开销，作为用户的用电习惯。
需要说明的是，以上充电习惯和用电习惯仅是举例说明，在本申请的基础上，还可以根据需求获取其他的信息作为功耗模型的输入。
2.功耗劣化量化
周期性地或在前台应用发生切换时，进行功耗劣化评估。功耗劣化评估的方式是计算当前应用对本时间片结束时待机时长的影响程度。
例如，在15:00，用户开始玩G游戏，该用户的充电习惯显示用户可能在21:00开始充电，当前电量剩余3000mA。功耗模型需要预留20％的电量避免进入省电模式，因此当前可分配电量为2800mA。15:00–18:00、18:00-21:00两个时间片中，用户历史用电比例为3:7，则两个时间片分别分配840mA和1960mA。当用户15:50结束游戏后，发现游戏消耗了300mA电量，按照用电习惯预估当前时间片内还要消耗800mA，则当前时间片功耗劣化程度为1960-(2800-300-800)/1960＝0.1326。功耗劣化触发功耗报警，将劣化报警发送给融合决策模块。
流程四.热控模型
出厂前，手机中会预置两部分只与器件型号、整机设备布局有关的信息：
a.热区器件的温升函数F i(R i,t)，表示档位在R i、持续时间t之后该器件产生的温升；
b.整机散热能力曲线。
在手机使用过程中，在不充电状态时通过卡尔曼滤波的方式对环境温度进行估计，具体方法可参考前述图7A和图7B。
在手机使用过程中，定时对热区温度进行采集，并根据采集的热区温度和热区的温升预测函数计算壳温的预测值。然后根据壳温的预测和环境温度产生温升预警。
例如，当前环境温度为26度。用户在玩G游戏的过程中，CPU高负荷高频点运行。通过各热区的温度采集和热区温度预测，发现当前壳温38度，将在5分钟内超过42度，短期内的最高温度(42度)超过了阈值(假设阈值是40度)，则产生温升预警。
报送到融合调度中心的是量化后的温度劣化结果，将类似<短期温升，当前38度，5分钟内超过42度>的预警信息转化为一个温度劣化值。这种转换可以是按公式进行，也可以根据规则进行。例如，将温度分为几个档位：38,42,45,48,53，在5分钟内温升超过对应档位分别对应温度劣化值0.2,0.3,0.5,0.8,1.2，十分钟内温升超过对应档位对应温度劣化值0.1,0.2,0.3,0.4,0.6。本实施例中温度劣化值输出为0.3。
主导热区的发现是在壳温计算的过程中进行的，将对壳温升高贡献最大的热区作为主 导热区(例如GPU)，发送到融合调度中心，供决策使用。
流程五.融合调度
当有性能劣化告警、功耗劣化告警或者温升告警中的一个或者多个事件出现时，会触发融合调度决策，即寻找最大化用户体验R的调度策略，请参考前述公式(9)。具体的α i和S i通过如下方式获得。
● 性能、功耗、热在当前应用场景下的权重α i
每个用户体验维度的权重系数α针对不同的应用场景、不同的用户均有所差异，可以由以下多个因素综合决定：
α＝γ(scene,time_domain,deterioration)   (10)
其中，γ(.)是权重计算公式，根据多个因素计算出三个维度各自的权重系数，多个因素包括：
a)scene是终端应用场景，例如游戏场景、视频场景、阅读场景等。不同场景下各个维度重要性有差异，例如游戏场景下，性能维度的权重要高于功耗和热维度的权重；本实施例中是：<节假日，晚上，家里，G游戏，团战，60帧>；
b)time_domain是该体验维度的作用时域，例如是秒级、分钟级别、小时级别，体验维度的作用时域越小则相对的权重会越大；
c)deterioration是该维度的劣化程度，劣化程度越严重则该维度的权重越大。
举例而言，在<游戏场景，在家>的条件下，当<性能劣化，功耗劣化，温升劣化>为<0.3,0.3095,0.3>的时候，最终计算的<性能维度权重，功耗维度的权重，温升维度的权重>为<0.5,0.3,0.2>。α的计算公式γ可以是预置的待查询数据，也可以是根据公式实时计算的数据。
● 性能、功耗、热的体验(S i)的量化
性能维度的用户体验：根据前台用户体验和后台用户体验来衡量。其中，前台用户体验由目标帧率、丢帧率、低帧率占比和流畅度来决定；后台用户体验由后台应用在后续使用的热启动比例来衡量。性能维度的用户体验的量化方式为平均帧率、低帧率占比、帧率抖动三个指标的满足程度，计算方式为1减去性能劣化程度1-ΔT，即性能模型中的性能劣化越小则用户体验越好，最大为1。
功耗维度的用户体验：根据待机时长来衡量。当待机时长未达到设置的上限时，待机时长越大则功耗维度的用户体验越佳；达到上限后，待机时长的增加不影响功耗维度的用户体验。例如，待机时长的上限可以为用户充电周期的1.2倍。功耗维度的用户体验量化方式为预测的待机时长和待机时长的上限的比值，最大为1。
温升维度的用户体验：当后续温度不会超过阈值时(即用户感知手机不发烫)，温升维 度的用户体验根据温升速度的最大值和方差来衡量，最大值越小且方差越小则温升体验最佳(控制温升速度以最平缓的速度上升)。热维度的用户体验量化方式为1减去温升劣化程度。最大为1。温升劣化程度的量化方式见“流程四.热控模型”。
● 最优调度策略搜索
图10为一种最优调度策略的确定过程的示例。如图所示，在可用的管控策略空间中寻找最优的策略的方法可以分为以下几步。
S901、借助性能瓶颈、主导热区来确定策略搜索的方向，针对性地选择出候选的资源调度策略集合。
性能瓶颈和主导热区分别由性能模型和热控模型所提供，能帮助减少策略的搜索空间，减少候选的资源调度策略的数目。例如在G游戏场景，假设G游戏的性能瓶颈是DDR，当发生性能劣化时应提升DDR资源来提升性能体验；假设发现主导热区是GPU，可针对性地降低GPU的频点来实现快速地控制温升。更具体地分析，当多个维度之间存在冲突时，例如主导热区和性能瓶颈都是大核资源，那么大核资源频点上升和下降都是可能的搜索方向，由后面的用户整体体验来决定。
除了性能瓶颈和主导热区之外，资源功耗表信息也可以作为筛选资源调度策略的输入。资源功耗表信息由功耗模型提供。资源功耗表中包括多种资源及其对应的耗电信息，例如，计算资源，如CPU和GPU等在什么档位，单位时间耗多少电；外设，如屏幕、声音等在什么档位，单位时间耗多少电；不同应用在后台单位时间平均耗多少电。
S902、针对每种候选的资源调度策略，结合性能、功耗、温度这三个维度的用户体验评估方法，来预估这三个维度分别的用户体验。各维度的用户体验评估方法详见前一部分“性能、功耗、热的体验(S i)的量化”。
S903、根据各个维度在当前应用场景下的权重系数，计算每种资源调度策略的整体用户体验，并进行排序选择得分最高的策略，作为最优的资源调度策略。计算公式为R＝∑α i*S i(相当于公式(9))，其中权重系数的衡量方法详见前一部分“性能、功耗、热在当前应用场景下的权重α i”。
S904、下发并执行最优的资源调度策略，并通过实时反馈(S905)来获取采取该策略后的各个体验维度(包括性能、功耗、温度)的真实变化情况，并借此来调整上述步骤中的计算参数。例如，假设根据之前的计算最优策略实施后性能维度不劣化，但策略执行后实时监控发现还存在丢帧导致性能体验有所下降，则反馈机制会将信息反馈到性能维度的用户体验评估计算处，更新各个参数的权重。进一步的，也会将信息反馈到性能模型的性能瓶颈分析处，重新学习该场景下的性能瓶颈。
需要说明的是，在以上实施例中说明了一些数据量化或数据处理方式，都是为了后续的数学模型能够顺利计算，但是本申请并不局限于这些实施例。在具体实现过程中，本领域技术人员容易根据本申请提出的方法想到其他的数学处理方法。另外，具体计算过程中可能涉及到误差。
图11为一种智能手机的设备结构示意图。图1为智能手机系统架构的分层示意图，而图11尽可能地从硬件角度描绘手机的结构。图1中的硬件130包括的硬件及其连接关系可以参考图11。图1中操作系统120和应用110，以软件程序的形式存储在图11所示的存储器870内部。
如图11所示，该手机包括通信模块810、传感器820、用户输入模块830、输出模块840、处理器850、音视频输入模块860、存储器870以及电源880。进一步的，本实施例提供的手机还可以包括神经网络处理单元(neural-network processing unit,NPU)890。
通信模块810可以包括至少一个能使该计算机系统与通信系统或其他计算机系统之间进行通信的模块。例如，通信模块810可以包括有线网络接口，广播接收模块、移动通信模块、无线因特网模块、局域通信模块和位置(或定位)信息模块等其中的一个或多个。这多种模块均在现有技术中有多种实现，本申请不一一描述。
传感器820可以感测系统的当前状态，诸如打开/闭合状态、位置、与用户是否有接触、方向、和加速/减速，并且传感器820可以生成用于控制系统的操作的感测信号。
用户输入模块830，用于接收输入的数字信息、字符信息或接触式触摸操作/非接触式手势，以及接收与系统的用户设置以及功能控制有关的信号输入等。用户输入模块830包括触控面板和/或其他输入设备。
输出模块840包括显示面板，用于显示由用户输入的信息、提供给用户的信息或系统的各种菜单界面等。可选的，可以采用液晶显示器(liquid crystal display，LCD)或有机发光二极管(organic light-emitting diode,OLED)等形式来配置显示面板。在其他一些实施例中，触控面板可覆盖显示面板上，形成触摸显示屏。另外，输出模块840还可以包括音频输出模块、告警器以及触觉模块等。
音视频输入模块860，用于输入音频信号或视频信号。音视频输入模块860可以包括摄像头和麦克风。
电源880可以在处理器850的控制下接收外部电力和内部电力，并且提供系统的各个组件的操作所需的电力。
虽然图示中将处理器850描绘成一个组件的样子，但850可以表示多个处理器，例如，处理器850可以包括一个中央处理器(central processing unit,CPU)和一个图形处理器(graphic processing unit,GPU)。CPU一般具有多个核，属于多核处理器。这多个核可以集成在同一块芯片上，也可以各自为独立的芯片。
存储器870存储计算机程序，该计算机程序包括操作系统程序872和应用程序871等。
典型的操作系统如微软公司的
Figure PCTCN2019104292-appb-000019
苹果公司的
Figure PCTCN2019104292-appb-000020
谷歌公司开发的基于
Figure PCTCN2019104292-appb-000021
的安卓
Figure PCTCN2019104292-appb-000022
系统等。本申请前述任意一个实施例提供的方法可以通过软件的方式实现，可以认为是操作系统程序872的具体实现，作为操作系统程序872的一部分部署在智能手机上，当处理器850读取操作系统程序872并运行之后，本申请提供的方法操作系统程序872的一个功能可适时运行起来。具体的，本申请提供 的功能可以默认开启，也可以在用户的指令下开启。
需要说明的是，操作系统程序872相当于图1中的操作系统120；应用程序871相当于图1中的应用110。
存储器870可以是以下类型中的一种或多种：闪速(flash)存储器、硬盘类型存储器、微型多媒体卡型存储器、卡式存储器(例如SD或XD存储器)、随机存取存储器(random access memory,RAM)(例如前述实施例中提到的DDR SDRAM)、只读存储器(read only memory,ROM)、电可擦除可编程只读存储器(electrically erasable programmable read-only memory，EEPROM)、可编程只读存储器(programmable ROM,PROM)、回滚保护存储块(replay protected memory block,RPMB)、磁存储器、磁盘或光盘。
在其他一些实施例中，存储器870也可以是因特网上的网络存储设备，系统可以对在因特网上的存储器870执行更新或读取等操作。
处理器850可以指示一个或多个处理单元，例如，处理器850可以包括一个或多个中央处理器，或者包括一个中央处理器和一个图形处理器，或者包括一个应用处理器和一个协处理器(例如微控制单元或神经网络处理器)。当处理器850包括多个处理单元时，这多个处理单元可以集成在同一块芯片上，也可以各自为独立的芯片。一个处理器可以包括一个或多个物理核。
存储器870还存储有除计算机程序之外的其他数据873。
NPU 890作为协处理器挂载到主处理器850上，用于执行主处理器850给它分配的任务。本申请实施例中，NPU890可以在主处理器850的调度下实现管控决策过程中涉及的部分复杂算法。在其它一些实施例中，NPU890也可以作为主处理器执行本申请提供的方法全部流程，或调用它的协处理器共同实现本申请提供的方法。
需要说明的是，图中示出的各个模块的连接关系仅为一种示例，本申请任意实施例提供的方法也可以应用在其它连接方式的智能手机中，例如所有模块或部分模块通过总线连接以实现通信。
图12为一种NPU900的结构示意图，该NPU900相当于图11中的NPU890。如图所示，NPU900与主处理器和外部存储器相连，这里的主处理器相当于图11中的处理器850。NPU900的核心部分为运算电路903，通过控制器904控制运算电路903提取存储器中的数据并进行数学运算。
在一些实现中，运算电路903内部包括多个处理引擎(process engine,PE)。在一些实现中，运算电路903是二维脉动阵列。运算电路903还可以是一维脉动阵列或者能够执行例如乘法和加法这样的数学运算的其它电子线路。在另一些实现中，运算电路903是通用的矩阵处理器。
举例来说，假设有输入矩阵A，权重矩阵B，输出矩阵C。运算电路903从权重存储器902中取矩阵B相应的数据，并缓存在运算电路903的每一个PE上。运算电路903从输入存储器901中取矩阵A数据与矩阵B进行矩阵运算，得到的矩阵的部分结果或最终结果，保存在累加器(accumulator)908中。
统一存储器906用于存放输入数据以及输出数据。权重数据直接通过存储单元访问控制器905(例如direct memory access controller，DMAC)被搬运到权重存储器902中。输 入数据也通过存储单元访问控制器905被搬运到统一存储器906中。
总线接口单元910(bus interface unit，BIU)用于AXI(advanced extensible interface)总线与存储单元访问控制器905和取指存储器909(instruction fetch buffer)的交互。
总线接口单元910用于取指存储器909从外部存储器获取指令，还用于存储单元访问控制器905从外部存储器获取输入矩阵A或者权重矩阵B的原数据。
存储单元访问控制器905主要用于将外部存储器中的输入数据搬运到统一存储器906或将权重数据搬运到权重存储器902中或将输入数据数据搬运到输入存储器901中。
向量计算单元907通常包括多个运算处理单元，在需要的情况下，对运算电路903的输出做进一步处理，如向量乘、向量加、指数运算、对数运算、和/或大小比较等等。
在一些实现中，向量计算单元907能将经处理的向量存储到统一存储器906中。例如，向量计算单元907可以将非线性函数应用到运算电路903的输出，例如累加值的向量，用以生成激活值。在一些实现中，向量计算单元907生成归一化的值、合并值，或二者均有。在一些实现中，经处理的向量能够用作运算电路903的激活输入。
与控制器904连接的取指存储器909用于存储控制器904使用的指令。
统一存储器906，输入存储器901，权重存储器902以及取指存储器909均为On-Chip存储器。图中的外部存储器与该NPU硬件架构独立。
本实施例提供的方法中存在多个模型和多种算法，其中涉及到的复杂算法，例如搜索最优调度策略的算法可以由NPU900执行，然后将执行结果返回给主处理器，从而提高方法的运行效率。
需要说明的是，NPU仅是举例，本申请提供的方案中的部分算法可以由其他类型的专用处理器处理。由于专用处理器在处理特定算法上的处理效率高于普通处理器，所以利用专用处理器处理方案中的部分算法，能够有效提升方案的整体运行效率，从而尽快找到合适的资源管控策略，并实现该资源管控策略。
需要说明的是，本实施例提供的方法也可以应用于非终端的计算机设备，例如云端服务器。
需要说明的是，以上实施例多以人脸识别方案为例介绍，但本申请提出的方法显然可以应用于除人脸识别之外的其它方案，本领域技术人员根据本申请提供的实现方式容易想到其它方案的类似实现方式。
需要说明的是，前述实施例中提出模块或单元的划分仅作为一种示例性的示出，所描述的各个模块的功能仅是举例说明，本申请并不以此为限。本领域普通技术人员可以根据需求合并其中两个或更多模块的功能，或者将一个模块的功能拆分从而获得更多更细粒度的模块，以及其他变形方式。
以上描述的各个实施例之间相同或相似的部分可相互参考。本申请中的“多个”若无特殊说明，指两个或两个以上，或“至少两个”。本申请中的“A/B”包括三种情况：“A”、“B”和“A和B”。
以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明的模块可以是或者也可以不是物理上分开的，作为模块显示的部件可以是或者也可以不是物理模块，即可以位于一个地方，或者也可以分布到多个网络模块上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。另外，本申请提供的装置实施例附图中，模块之间的连接关系表示它们之间具有通信连接，具体可以实现为一条或多条通信总线或信号线。本领域普通技术人员在不付出创造性劳动的情况下，即可以理解并实施。
以上所述，仅为本申请的一些具体实施方式，但本申请的保护范围并不局限于此。

Claims (33)

  1. 一种资源调度方法，其特征在于，包括：
    确定设备当前的应用场景；
    基于性能、功耗和温度在所述应用场景下分别的权重确定资源管控策略；
    根据所述资源管控策略调度所述设备的资源。
  2. 根据权利要求1所述的方法，其特征在于，所述基于性能、功耗和温度在所述应用场景下分别的权重确定资源管控策略，包括：
    当性能劣化程度、功耗劣化程度以及温度劣化程度中的任意一个或多个达到预设条件时，基于性能、功耗和温度在所述应用场景下分别的权重确定资源管控策略；
    其中：所述性能劣化程度用于指示所述设备的性能参数值与目标性能指标之间的差距，所述目标性能指标为满足所述应用场景下的性能体验的一个或多个性能参数的值；所述功耗劣化程度用于指示预测的可用电量与最低可用电量之间的差距，所述温度劣化程度用于指示预测的外壳温度和温度阈值之间的差距。
  3. 根据权利要求2所述的方法，其特征在于，所述目标性能指标与所述应用场景具有对应关系。
  4. 根据权利要求3所述的方法，其特征在于，还包括：对采集到的历史场景数据进行聚类以获得多个应用场景；为每个应用场景设置对应的目标性能指标。
  5. 根据权利要求2-4任意一项所述的方法，其特征在于，所述设备的性能参数值是根据所述应用场景和所述设备当前的系统状态预测得到的，所述系统状态指示所述设备当前的资源使用情况。
  6. 根据权利要求5所述的方法，其特征在于，所述系统状态包括资源供给信息和系统负载信息。
  7. 根据权利要求2-4任意一项所述的方法，其特征在于，所述设备的性能参数值为采集得到的性能参数值。
  8. 根据权利要求2-7任意一项所述的方法，其特征在于，所述功耗劣化程度为基于所述设备的电量使用规律以及所述设备的剩余电量确定的，其中，所述电量使用规律反映所述设备历史的充电和耗电情况。
  9. 根据权利要求8所述的方法，其特征在于，所述预测的可用电量通过如下方式预测：
    根据所述电量使用规律和当前的前台应用预测电量消耗，并获得所述预测的可用电量，所述预测的可用电量为所述剩余电量与所述电量消耗的差。
  10. 根据权利要求9所述的方法，其特征在于，所述电量使用规律包括充电周期、用电规律和应用使用规律，其中所述充电周期代表历史上的充电间隔，所述用电规律代表历史上多个时刻以及每个时刻的剩余电量，所述应用使用规律代表历史上应用对电量的消耗信息；
    所述最低可用电量是根据所述充电周期和所述用电规律确定的；所述电量消耗是根据所述应用使用规律和所述前台应用确定的。
  11. 根据权利要求10所述的方法，其特征在于，所述充电周期为充电结束时刻到下次 充电开始时刻之间的时间段；所述用电规律包括历史上的多个时刻以及每个时刻对应的剩余电量；所述应用使用规律包括反映应用的使用时长和使用功耗的信息。
  12. 根据权利要求2-11任意一项所述的方法，其特征在于，所述外壳温度是根据所述设备的当前外壳温度、所述设备内多个器件各自的温度升高预测值、以及每个器件温度的权重预测的。
  13. 根据权利要求12所述的方法，其特征在于，所述温度升高预测值根据器件的温升函数和所述器件的当前资源配置确定。
  14. 根据权利要求2-13任意一项所述的方法，其特征在于，所述温度阈值是根据所述设备当前所处的环境的温度确定的。
  15. 根据权利要求2-14任意一项所述的方法，其特征在于，所述温度劣化程度包括预测的一段时间内的外壳温度的最大值和第一阈值的第一差距以及温度波动值与第二阈值的第二差距，其中所述温度波动值为所述最大值与该段时间内的最小值之差。
  16. 根据权利要求1-15任意一项所述的方法，其特征在于，所述资源管控策略包括以下三项中的任意一项或多项：计算资源调度策略、应用管控策略以及外设管控策略。
  17. 根据权利要求1-16任意一项所述的方法，其特征在于，所述资源管控策略为使得所述应用场景下用户整体体验最优的资源管控策略，其中所述用户整体体验根据性能、功耗和温度三者的用户体验以及三者分别在所述应用场景下的所述权重获得。
  18. 根据权利要求1-17任意一项所述的方法，其特征在于，所述资源管控策略的确定过程利用机器学习或动态规划的方法。
  19. 根据权利要求2-18任意一项所述的方法，其特征在于，所述目标性能指标包括以下四项参数中的任意一项或多项的值：目标帧率、丢帧率、低帧率占比和流畅度；其中，所述低帧率占比为低于目标帧率的帧数占总帧数的百分比。
  20. 根据权利要求1-19任意一项所述的方法，其特征在于，所述应用场景包括应用内的使用场景。
  21. 根据权利要求1-20任意一项所述的方法，其特征在于，所述确定资源管控策略包括：
    根据设备的性能瓶颈或主导热区确定候选的资源管控策略，所述候选的资源管控策略能够提升所述性能瓶颈或降低所述主导热区的温度；
    从所述候选的资源管控策略中，基于所述性能、功耗和温度分别的权重搜索出所述资源管控策略。
  22. 根据权利要求1-21任意一项所述的方法，其特征在于，确定所述性能、功耗和温度各自的权重的因素包括以下任意一个或多个：所述设备的应用场景、所述性能、功耗和温度各自的劣化程度、以及所述性能、功耗和温度劣化判断的作用时域。
  23. 根据权利要求1-22任意一项所述的方法，其特征在于，所述基于性能、功耗和温度在所述应用场景下分别的权重确定资源管控策略包括：
    根据性能、功耗和温度三者各自的用户体验和各自的权重确定用户整体体验；
    根据用户整体体验确定资源管控策略。
  24. 根据权利要求23所述的方法，其特征在于，所述根据性能、功耗和温度三者各自 的用户体验满足程度和各自的权重确定用户整体体验包括：
    根据性能劣化程度、功耗劣化程度、温度劣化程度分别确定所述性能的体验值、所述功耗的体验值和所述温度的体验值，其中，所述体验值用于指示用户体验满足程度，所述劣化程度用于指示目标指标与实际指标的差距，劣化程度越高，对应的体验值越低；
    根据所述性能、功耗和温度的三个体验值以及三者在所述应用场景下分别的权重确定用户整体体验。
  25. 一种资源调度方法，其特征在于，包括：
    确定设备当前的应用场景；
    根据所述应用场景，融合性能、功耗和温度三者各自的用户体验以获得用户整体体验；
    根据所述用户整体体验确定并执行资源管控策略。
  26. 根据权利要求25所述的方法，其特征在于，所述根据所述应用场景，融合性能、功耗和温度三者各自的用户体验以获得用户整体体验包括：
    根据性能、功耗和温度三者各自的用户体验以及三者在所述应用场景下分别的权重获得用户整体体验。
  27. 根据权利要求25或26所述的方法，其特征在于，所述应用场景包括应用内的使用场景。
  28. 根据权利要求25-27任意一项所述的方法，其特征在于，所述确定资源管控策略包括：
    根据设备的性能瓶颈或主导热区确定候选的资源管控策略，所述候选的资源管控策略能够提升所述性能瓶颈或降低所述主导热区的温度；
    从所述候选的资源管控策略中，根据所述用户整体体验确定搜索出所述资源管控策略，其中，所述资源管控策略满足所述用户整体体验。
  29. 根据权利要求26-28任意一项所述的方法，其特征在于，确定所述性能、功耗和温度各自的权重的因素包括以下任意一个或多个：所述设备的应用场景、所述性能、功耗和温度各自的劣化程度、以及所述性能、功耗和温度劣化判断的作用时域。
  30. 一种计算机设备，其特征在于，所述计算机设备包括处理器和存储器，其中：所述存储器用于存储程序；所述处理器用于读取所述程序并实现如权利要求1-29中任意一项所述的方法。
  31. 根据权利要求30所述的计算机设备，其特征在于，所述计算机设备为终端设备。
  32. 一种计算机存储介质，其特征在于，包括程序，所述程序在被一个或多个处理单元执行时实现如权利要求1-29中任意一项所述的方法。
  33. 一种计算机程序产品，其特征在于，包括程序，所述程序在被一个或多个处理单元执行时实现如权利要求1-29中任意一项所述的方法。
PCT/CN2019/104292 2018-10-15 2019-09-04 资源调度方法和计算机设备 Ceased WO2020078135A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP19872996.4A EP3848776A4 (en) 2018-10-15 2019-09-04 RESOURCE PLANNING PROCESS AND COMPUTER DEVICE
US17/231,268 US20210232201A1 (en) 2018-10-15 2021-04-15 Resource scheduling method and computer device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811198978.8 2018-10-15
CN201811198978.8A CN109960395B (zh) 2018-10-15 2018-10-15 资源调度方法和计算机设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/231,268 Continuation US20210232201A1 (en) 2018-10-15 2021-04-15 Resource scheduling method and computer device

Publications (1)

Publication Number Publication Date
WO2020078135A1 true WO2020078135A1 (zh) 2020-04-23

Family

ID=67023222

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/104292 Ceased WO2020078135A1 (zh) 2018-10-15 2019-09-04 资源调度方法和计算机设备

Country Status (4)

Country Link
US (1) US20210232201A1 (zh)
EP (1) EP3848776A4 (zh)
CN (1) CN109960395B (zh)
WO (1) WO2020078135A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220261305A1 (en) * 2021-02-18 2022-08-18 Ajou University Industry-Academic Cooperaiton Foundation Apparatus and method for optimizing reliability of satellite system considering both hard error stability and soft error stability
CN116777730A (zh) * 2023-08-25 2023-09-19 湖南马栏山视频先进技术研究院有限公司 一种基于资源调度的gpu效能提高方法
CN118690096A (zh) * 2024-08-28 2024-09-24 深圳开鸿数字产业发展有限公司 一种基于资源调度的调频方法、系统及终端
CN120216176A (zh) * 2025-03-10 2025-06-27 深圳市神速信息技术有限公司 一种服务器资源调度方法及服务器

Families Citing this family (44)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109960395B (zh) * 2018-10-15 2021-06-08 华为技术有限公司 资源调度方法和计算机设备
CN110928587A (zh) * 2019-11-14 2020-03-27 联想(北京)有限公司 控制方法和控制装置
CN110955580B (zh) * 2019-12-09 2023-10-03 Oppo广东移动通信有限公司 壳体温度的获取方法、装置、存储介质和电子设备
CN113537264B (zh) * 2020-04-21 2022-09-20 阿里巴巴集团控股有限公司 空间的应用状态检测方法、管理方法、装置及设备
CN111954286B (zh) * 2020-08-06 2022-04-22 捷开通讯(深圳)有限公司 一种功率调节方法、装置、存储介质及终端
CN111966475B (zh) * 2020-08-13 2023-03-24 Oppo（重庆）智能科技有限公司 应用程序切换方法及终端设备、计算机可读存储介质
CN112068688A (zh) * 2020-09-09 2020-12-11 北京字节跳动网络技术有限公司 省电方法、装置、电子设备及计算机可读介质
CN112162625B (zh) * 2020-09-25 2024-10-25 上海闻泰信息技术有限公司 电子设备的功耗控制方法、装置、存储介质和终端
CN112587920B (zh) * 2020-12-01 2024-09-24 Oppo（重庆）智能科技有限公司 设备控制方法、装置、电子设备及存储介质
CN112423324B (zh) * 2021-01-22 2021-04-30 深圳市科思科技股份有限公司 无线智能决策通信方法、装置和系统
JP7333839B2 (ja) 2021-01-22 2023-08-25 深▲せん▼市科思科技股▲ふん▼有限公司 無線インテリジェント意思決定通信方法、装置及びシステム
CN113795030A (zh) * 2021-09-13 2021-12-14 广州点金石信息技术有限公司 一种网络性能优化方法、装置、移动终端和存储介质
CN114037271B (zh) * 2021-11-09 2025-05-02 国泰新点软件股份有限公司 警力资源分配方法及装置
CN114444853B (zh) * 2021-12-17 2025-07-25 北京达佳互联信息技术有限公司 一种性能控制方法、装置及存储介质
CN116366956A (zh) * 2021-12-24 2023-06-30 北京小米移动软件有限公司 相机功耗的调节方法及装置
TWI849735B (zh) * 2022-02-10 2024-07-21 聯發科技股份有限公司 用於平衡功率和性能的自我調整負載感知系統管理
CN115599513B (zh) * 2022-04-07 2023-11-03 荣耀终端有限公司 资源调度方法及电子设备
US12066882B2 (en) * 2022-04-12 2024-08-20 Meta Platforms Technologies, Llc Systems and methods for managing power consumption of device subsystems
CN114721921A (zh) * 2022-04-25 2022-07-08 Oppo广东移动通信有限公司 用户感受评估方法、装置、终端设备及存储介质
CN114788955B (zh) * 2022-04-27 2025-01-28 Oppo广东移动通信有限公司 性能参数调整方法、装置、电子设备、存储介质及产品
CN117873693A (zh) * 2022-05-16 2024-04-12 荣耀终端有限公司 资源调度方法、电子设备及存储介质
CN116027880B (zh) * 2022-05-16 2023-11-24 荣耀终端有限公司 资源调度方法和电子设备
CN116028207B (zh) * 2022-05-16 2024-04-12 荣耀终端有限公司 调度策略确定方法、装置、设备和存储介质
CN116027878B (zh) * 2022-05-16 2023-10-20 荣耀终端有限公司 功耗调整方法和电子设备
CN116028208B (zh) * 2022-05-16 2023-10-20 荣耀终端有限公司 系统负载确定方法、装置、设备和存储介质
CN116089055B (zh) * 2022-05-16 2024-04-02 荣耀终端有限公司 资源调度方法和装置
CN116054297B (zh) * 2022-05-30 2024-03-29 荣耀终端有限公司 一种充电方法、装置及电子设备
CN117369620A (zh) * 2022-06-30 2024-01-09 北京小米移动软件有限公司 电子设备的控制方法、装置、电子设备和存储介质
CN117393882A (zh) * 2022-07-04 2024-01-12 北京小米移动软件有限公司 剩余使用时长的预测方法、装置、电子设备及存储介质
US12113955B2 (en) 2022-09-12 2024-10-08 Apple Inc. Head-mounted electronic device with adjustable frame rate
CN115271000B (zh) * 2022-09-26 2022-12-02 国网山西省电力公司太原供电公司 一种用于电缆隧道的状态监测方法及系统
CN117784837A (zh) * 2022-09-27 2024-03-29 蔚来移动科技有限公司 基于操作系统的温度控制方法、系统、装置及介质
CN117950935B (zh) * 2022-10-31 2025-01-03 荣耀终端有限公司 一种性能调控方法及电子设备
US12598484B2 (en) * 2022-12-02 2026-04-07 Samsung Electronics Co., Ltd. Traffic aware UE temperature management
US20240201776A1 (en) * 2022-12-16 2024-06-20 Apple Inc. Electronic Device with Centralized User Experience Manager
CN115639903B (zh) * 2022-12-23 2023-04-07 湖南博盛芯微电子科技有限公司 基于飞腾平台的计算机主板及计算机功耗控制方法
TWI872456B (zh) * 2023-02-14 2025-02-11 華碩電腦股份有限公司 電子裝置及其效能優化方法
EP4654012A4 (en) * 2023-03-16 2026-03-04 Huawei Tech Co Ltd Computing power resource scheduling method and apparatus
US12598264B2 (en) * 2023-05-31 2026-04-07 Apple Inc. Color correction for electronic device with immersive viewing
US12548491B2 (en) 2023-10-04 2026-02-10 Samsung Electronics Co., Ltd. Device and methods for managing frame drops for enhanced user device experience
WO2025075268A1 (en) * 2023-10-04 2025-04-10 Samsung Electronics Co., Ltd. Device and methods for managing frame drops for enhanced user device experience
CN119862024A (zh) * 2023-10-20 2025-04-22 蔚来移动科技有限公司 设备资源调度方法、控制装置及存储介质
CN117806831B (zh) * 2024-02-23 2024-07-19 荣耀终端有限公司 参数配置方法及相关设备
CN120974087B (zh) * 2025-08-27 2026-03-24 沈阳科来沃电气技术有限公司 基于平衡畸变的逆变器电流归一化方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150186184A1 (en) * 2013-12-26 2015-07-02 Electronics And Telecommunications Research Institute Apparatus and method for optimizing system performance of multi-core system
CN106774786A (zh) * 2016-11-22 2017-05-31 珠海市魅族科技有限公司 一种功耗控制方法以及装置
CN107515787A (zh) * 2017-08-31 2017-12-26 广东欧珀移动通信有限公司 资源配置方法及相关产品
CN107577533A (zh) * 2017-08-31 2018-01-12 广东欧珀移动通信有限公司 资源配置方法及相关产品
CN109960395A (zh) * 2018-10-15 2019-07-02 华为技术有限公司 资源调度方法和计算机设备

Family Cites Families (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7051221B2 (en) * 2003-04-28 2006-05-23 International Business Machines Corporation Performance throttling for temperature reduction in a microprocessor
US8237386B2 (en) * 2003-08-15 2012-08-07 Apple Inc. Methods and apparatuses for operating a data processing system
US8452992B2 (en) * 2010-06-29 2013-05-28 Stmicroelectronics S.R.L. Power management using constraints in multi-dimensional parameter space
US8527994B2 (en) * 2011-02-10 2013-09-03 International Business Machines Corporation Guarded, multi-metric resource control for safe and efficient microprocessor management
CN105247486B (zh) * 2013-05-23 2019-05-21 瑞萨电子株式会社 多cpu系统及多cpu系统的调整方法
US20140362518A1 (en) * 2013-06-10 2014-12-11 Wei Huang Thermal management of a portable computing device
US9703355B2 (en) * 2013-08-28 2017-07-11 Qualcomm Incorporated Method, devices and systems for dynamic multimedia data flow control for thermal power budgeting
CN105659189A (zh) * 2013-11-29 2016-06-08 联发科技股份有限公司 用于对便携式装置上的系统进行功率调节的方法和控制器、相应的便携式装置和相应的计算机程序产品
GB2540804B (en) * 2015-07-29 2018-03-07 Samsung Electronics Co Ltd Hardware power management apparatus and methods
US20170228849A1 (en) * 2016-02-05 2017-08-10 Mediatek Inc. Apparatuses and methods for activity-based resource management, and storage medium thereof
US9934048B2 (en) * 2016-03-29 2018-04-03 Intel Corporation Systems, methods and devices for dynamic power management of devices using game theory
CN106095063A (zh) * 2016-06-15 2016-11-09 乐视控股（北京）有限公司 移动终端的功耗管理方法和功耗管理系统
CN106569580A (zh) * 2016-10-21 2017-04-19 哈尔滨工业大学深圳研究生院 智能终端的动态节能方法及系统
CN108021219A (zh) * 2016-11-03 2018-05-11 华为技术有限公司 电子设备控制方法、装置及电子设备
US20180136709A1 (en) * 2016-11-16 2018-05-17 Microsoft Technology Licensing, Llc Dynamic External Power Resource Selection
US11656666B2 (en) * 2016-11-16 2023-05-23 Microsoft Technology Licensing, Llc Dynamic power source selection, charging, and discharging
US20180157315A1 (en) * 2016-12-01 2018-06-07 Qualcomm Incorporated System and method for proactive power and performance management of a workload in a portable computing device
US20190196567A1 (en) * 2016-12-08 2019-06-27 Shenzhen Royole Technologies Co. Ltd. Electronic device and sleep control method thereof
CN106970696B (zh) * 2017-03-24 2020-04-24 联想(北京)有限公司 一种电子设备管理方法及电子设备
CN107370877B (zh) * 2017-06-30 2020-06-02 Oppo广东移动通信有限公司 应用程序的推送方法、装置、存储介质及电子设备
CN107832142B (zh) * 2017-10-11 2020-06-09 Oppo广东移动通信有限公司 应用程序的资源配置方法及设备
CN107861816B (zh) * 2017-10-31 2022-10-28 Oppo广东移动通信有限公司 资源配置方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150186184A1 (en) * 2013-12-26 2015-07-02 Electronics And Telecommunications Research Institute Apparatus and method for optimizing system performance of multi-core system
CN106774786A (zh) * 2016-11-22 2017-05-31 珠海市魅族科技有限公司 一种功耗控制方法以及装置
CN107515787A (zh) * 2017-08-31 2017-12-26 广东欧珀移动通信有限公司 资源配置方法及相关产品
CN107577533A (zh) * 2017-08-31 2018-01-12 广东欧珀移动通信有限公司 资源配置方法及相关产品
CN109960395A (zh) * 2018-10-15 2019-07-02 华为技术有限公司 资源调度方法和计算机设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3848776A4 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220261305A1 (en) * 2021-02-18 2022-08-18 Ajou University Industry-Academic Cooperaiton Foundation Apparatus and method for optimizing reliability of satellite system considering both hard error stability and soft error stability
US11615011B2 (en) * 2021-02-18 2023-03-28 Ajou University Industry-Academic Cooperation Foundation Apparatus and method for optimizing reliability of satellite system considering both hard error stability and soft error stability
CN116777730A (zh) * 2023-08-25 2023-09-19 湖南马栏山视频先进技术研究院有限公司 一种基于资源调度的gpu效能提高方法
CN116777730B (zh) * 2023-08-25 2023-10-31 湖南马栏山视频先进技术研究院有限公司 一种基于资源调度的gpu效能提高方法
CN118690096A (zh) * 2024-08-28 2024-09-24 深圳开鸿数字产业发展有限公司 一种基于资源调度的调频方法、系统及终端
CN120216176A (zh) * 2025-03-10 2025-06-27 深圳市神速信息技术有限公司 一种服务器资源调度方法及服务器

Also Published As

Publication number Publication date
EP3848776A4 (en) 2021-11-10
EP3848776A1 (en) 2021-07-14
US20210232201A1 (en) 2021-07-29
CN109960395A (zh) 2019-07-02
CN109960395B (zh) 2021-06-08

Similar Documents

Publication Publication Date Title
CN109960395B (zh) 资源调度方法和计算机设备
US10554786B2 (en) Dynamic adjustment of mobile device based on peer event data
US10936358B2 (en) Initiating background updates based on user activity
US9462965B2 (en) Dynamic adjustment of mobile device based on system events
US9432839B2 (en) Dynamic adjustment of mobile device based on thermal conditions
US9256484B2 (en) Dynamic adjustment of mobile device based on user activity
US20150347205A1 (en) Dynamic Adjustment of Mobile Device Based on Adaptive Prediction of System Events
US20150351033A1 (en) Dynamic Adjustment of Mobile Device Based on Voter Feedback
CN113728294B (zh) 功耗控制与方案生成方法、设备、系统及存储介质
CN115202889B (zh) 计算资源调整方法及计算系统
WO2020063290A1 (zh) 一种基于用户习惯的电量控制方法及终端
CN107748697B (zh) 应用关闭方法、装置、存储介质及电子设备
CN112306406A (zh) 智能存储自动分级的方法、装置、存储介质及电子设备
CN118573589A (zh) 一种掌上银行网络流量预测方法和相关装置
US20230088429A1 (en) Processing device, processing method, and program
CN121067429A (zh) 一种通讯终端控制方法、电子设备、存储介质
CN121742593A (zh) 任务执行方法、装置、设备、介质及产品
CN120751472A (zh) 功率调整方法及相关设备
CN121682117A (zh) 一种基于充值熵的主动干预决策方法及系统
CN120280920A (zh) 一种用电供需协同智能控制方法、装置、设备以及介质
CN121900943A (zh) 一种基于Kubernetes的循环资源调度方法、系统、设备及介质
HK40036406B (zh) 信息推荐方法、装置和电子设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19872996

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019872996

Country of ref document: EP

Effective date: 20210407

NENP Non-entry into the national phase

Ref country code: DE