KR20180071056A - Apparatus and method for dynamic frequency scaling, and recording medium - Google Patents

Abstract

Disclosed are a dynamic frequency scaling apparatus, a dynamic frequency scaling method, and a recording medium. According to an embodiment of the present invention, the dynamic frequency scaling method includes: a learning step of calculating a target frequency for a processing device by determining a minimum frequency that satisfies a maximum performance index of an application program among frequencies allowable to the processing device executing the application program, and setting a target performance index range based on the maximum performance index; and a tracking step of scaling an operating frequency of the processing device based on the performance index, calculated during the execution of the application program, the target performance index range, and the target frequency. According to the embodiment of the present invention, it is unnecessary to perform a static analysis before the execution of the application program, and a high power gain is obtained by performing dynamic frequency scaling that satisfies performance requirements in real time.

Description

TECHNICAL FIELD [0001] The present invention relates to a dynamic frequency adjusting apparatus, a dynamic frequency adjusting apparatus, a dynamic frequency adjusting apparatus,

The present invention relates to a dynamic frequency adjusting apparatus, a dynamic frequency adjusting method, and a recording medium, and more particularly, to a dynamic frequency adjusting apparatus, a dynamic frequency adjusting method, and a recording medium for reducing power by adjusting an operating frequency of a processing apparatus, A dynamic frequency adjusting method, and a recording medium.

Like complex mobile games and augmented reality, applications are increasingly demanding high-end graphics in mobile environments. Since mobile devices use batteries with a limited energy capacity as a power source, it is very important to efficiently manage the power used in the devices while satisfying the demand for application quality of service.

One of the most efficient techniques for power management is dynamic voltage frequency scaling (DVFS), which dynamically adjusts the operating frequency of a central processing unit (CPU) or a graphics processing unit (GPU). Of the various DVFS techniques, the on-demand DVFS technique is most often used for GPU. This technique performs power management in such a manner that the utilization rate of the processing apparatus is used as an index and the operating frequency of the processing apparatus is changed when the utilization rate is outside the critical range. However, this technique has a disadvantage in that it can not guarantee frame rate (FPS) performance, which is a representative quality of service indicator of mobile games.

In order to solve the problem of the on-demand DVFS technique, DVFS techniques using FPS as an index have recently been proposed. These techniques lower the operating frequency when the current FPS is higher than the target FPS and perform the power management by raising the operating frequency when the current FPS is lower than the target FPS. In the case of these FPS-based techniques, the target FPS is set in advance and the power can not be efficiently managed for various applications. In addition, the conventional method has a disadvantage that the frequency scale value according to the FPS must be set through static regression analysis before execution of the application program. Also, in the case of a heterogeneous multiprocessing computer system in which both CPU clusters and GPUs are used, it is not easy to derive the relationship between the usage rate of processors (CPU, GPU) and FPS performance through offline regression analysis.

The present invention relates to a dynamic frequency control device, a dynamic frequency control method, and a recording medium, which can achieve high power gain by performing dynamic frequency control that satisfies performance requirements in real time without performing static analysis before execution of an application program The purpose is to provide.

In addition, the present invention can adaptively adjust the operation frequency of the processing units constituting the heterogeneous computer system based on the utilization rate and the performance index (for example, FPS) An apparatus, a dynamic frequency adjustment method, and a recording medium.

The problems to be solved by the present invention are not limited to the above-mentioned problems. Other technical subjects not mentioned will be apparent to those skilled in the art from the description below.

A dynamic frequency adjustment apparatus according to an aspect of the present invention determines a minimum frequency that satisfies a maximum performance index of an application program among allowable frequencies for a processing apparatus that executes an application program and calculates a target frequency for the processing apparatus Sets a target performance index range based on the maximum performance index, and adjusts the operating frequency of the processing apparatus based on the performance index calculated during execution of the application program, the target performance index range, and the target frequency.

Wherein the dynamic frequency adjustment device calculates the minimum frequency that satisfies the maximum performance index among the allowable frequencies for the processing device at the target frequency and calculates an upper limit value and a lower limit value of the target performance index range based on the maximum performance index A learning unit for determining a learning result; A tracking unit for monitoring whether the performance index calculated during execution of the application program is out of the target performance index range and increasing the operation frequency of the processing unit or requesting the learning unit to re-determine the target performance index range; And a verification unit for comparing the current performance index calculated for the application program with the target performance indicator range after the operating frequency of the processing unit is increased by the tracking unit and analyzing the cause of the decrease in the performance indicator .

Wherein the processing unit includes a plurality of processing units, wherein the learning unit determines a priority order for the plurality of processing units based on a usage rate of the plurality of processing units, sequentially orders the plurality of processing units The target frequencies can be calculated by determining the combination of the minimum frequencies satisfying the maximum performance index.

Wherein the tracking unit increases the operating frequency based on the performance indicator, the maximum performance indicator, and the target frequency when the performance indicator is smaller than the lower limit value of the target performance indicator range, And when the performance index is larger than the upper limit value of the target performance index range, after setting the maximum frequency among the frequencies as the operation frequency, the learning performance index range and the target frequency And when the performance index satisfies the target performance index range, compares the usage rate of the processing apparatus with a preset threshold value, and when the utilization rate is smaller than the threshold value, have.

Wherein the confirming unit compares the current performance indicator with a past performance indicator calculated before the operation frequency is increased when it is determined that the current performance indicator does not satisfy the target performance indicator range, The control unit may request the learning unit to reset the target performance index range after adjusting the operation frequency based on the current performance index.

According to another aspect of the present invention, a target frequency for the processing apparatus is determined by determining a minimum frequency that satisfies the maximum performance index of the application program among the allowable frequencies of the processing apparatus executing the application program, A learning step of setting a target performance indicator range based on the indicator; And a tracking step of adjusting the operating frequency of the processing apparatus based on the performance index calculated during execution of the application program, the target performance index range, and the target frequency.

Wherein the dynamic frequency adjustment method further comprises the step of comparing the current performance index calculated for the application program with the target performance index range after the operating frequency of the processing apparatus is increased in the tracking step, ; ≪ / RTI >

Wherein the tracking step increases the operating frequency based on the performance index, the maximum performance index and the target frequency when the performance index is smaller than the lower limit value of the target performance index range, Switching to the confirming step for discrimination; Setting the maximum frequency among the frequencies to the operating frequency of the processing device when the performance index is larger than the upper limit value of the target performance index range, and then returning the target performance index range and the target frequency to the learning step Switching; And when the performance index satisfies the target performance index range, comparing the usage rate of the processing device with a preset threshold value, and decreasing an operation frequency of the processing device when the utilization rate is less than the threshold value .

Wherein the confirming step compares the current performance indicator with the target performance indicator range so that when the current performance indicator does not satisfy the target performance indicator range, Comparing with the calculated past performance indicators; And adjusting the operating frequency based on the current performance index, the maximum performance index, and the target frequency when the current performance index is determined to be smaller than the past performance index, And switching to the learning step in order to perform the learning.

Wherein the learning step comprises the steps of: determining a priority order for the plurality of processing units based on utilization rates of the plurality of processing units constituting the processing apparatus; And calculating target frequencies by sequentially determining a combination of the minimum frequencies satisfying the maximum performance index for the plurality of processors in accordance with the priority order.

According to another aspect of the present invention, there is provided a computer-readable recording medium on which a program for executing the dynamic frequency adjustment method is recorded.

According to the embodiment of the present invention, it is possible to provide a dynamic frequency adjusting device capable of obtaining a high power gain by performing dynamic frequency adjustment that satisfies performance requirements in real time without performing static analysis before execution of an application program, An adjustment method and a recording medium are provided.

In addition, according to the embodiment of the present invention, the operation frequency of the processing units constituting the heterogeneous computer system is adaptively adjusted based on the utilization rate and the performance index (for example, FPS) to execute the high- A dynamic frequency adjusting method, and a recording medium are provided.

The effects of the present invention are not limited to the effects described above. Unless stated, the effects will be apparent to those skilled in the art from the description and the accompanying drawings.

1 is a block diagram of a computer system including a dynamic frequency adjusting apparatus according to an embodiment of the present invention.
2 is a configuration diagram of a dynamic frequency adjusting apparatus according to an embodiment of the present invention.
3 is an overall flowchart of a dynamic frequency adjustment method according to an embodiment of the present invention.
4 is a state diagram of step S100 of FIG.
5 is a time diagram showing a frame generation time point, an FPS update time point, and a frequency adjustment time point.
FIG. 6 is a state diagram illustrating the learning state shown in FIG. 4 in detail.
7 is a flowchart of a tracking step constituting a dynamic frequency adjustment method according to an embodiment of the present invention.
FIG. 8 is a flow chart of an identifying step for configuring a dynamic frequency adjusting method according to an embodiment of the present invention.
9 to 12 are graphs showing performance (power gain, FPS) of the dynamic frequency adjustment method according to the embodiment of the present invention.

Other advantages and features of the present invention and methods of achieving them will be apparent by referring to the embodiments described hereinafter in detail with reference to the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, and the present invention is only defined by the scope of the claims. Although not defined, all terms (including technical or scientific terms) used herein have the same meaning as commonly accepted by the generic art in the prior art to which this invention belongs. A general description of known configurations may be omitted so as not to obscure the gist of the present invention. In the drawings of the present invention, the same reference numerals are used as many as possible for the same or corresponding configurations.

The terminology used in this application is used only to describe a specific embodiment and is not intended to limit the invention. In the present application, the terms "comprises", "having", or "having" are intended to specify the presence of stated features, integers, steps, operations, components, Steps, operations, elements, parts, or combinations thereof, whether or not explicitly described or implied by the accompanying claims.

Used throughout this specification may refer to a hardware component such as, for example, software, FPGA or ASIC, as a unit for processing at least one function or operation. However, "to" is not meant to be limited to software or hardware. &Quot; to " may be configured to reside on an addressable storage medium and may be configured to play one or more processors.

As an example, the term '~' includes components such as software components, object-oriented software components, class components and task components, and processes, functions, attributes, procedures, Routines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. The functions provided by the components and components may be performed separately by a plurality of components and components, or may be integrated with other additional components.

The dynamic frequency adjustment apparatus according to an embodiment of the present invention determines a minimum frequency that satisfies a maximum performance index of an application program among frequencies applicable to a processing apparatus that executes an application program, and calculates a target frequency for the processing apparatus. The dynamic frequency control unit sets a target performance index range based on the maximum performance index. The dynamic frequency control device adjusts the operating frequency of the processing device adaptively to a performance index that is calculated in real time during execution of an application program based on the target performance index range and the target frequency.

According to the embodiment of the present invention, there is no need to perform a static analysis before execution of an application program, dynamic voltage frequency scaling (DVFS) is performed while satisfying performance requirements in real time, Can be obtained. In addition, the target frequencies may be sequentially determined for the processing units (for example, the CPU and the GPU) of the heterogeneous computer system to increase the power gain of the heterogeneous computer system.

In one embodiment, if the application is a 3D mobile game, the performance indicator may illustratively include a frame rate (FPS). However, the present invention can be applied to various application programs other than games, and other performance indicators other than the frame rate may be used depending on an application program or a system to which the present invention is applied.

For example, when the present invention is applied to an embedded system that processes image recognition, a performance index related to the number of recognized images per hour can be used, and when applied to a system that performs machine learning, It is possible. Hereinafter, the present invention will be described with reference to an embodiment in which a frame rate (FPS) is used as a performance index. However, the application program to which the present invention is applied and the kind of the performance index are not limited to these examples.

1 is a block diagram of a computer system including a dynamic frequency adjusting apparatus according to an embodiment of the present invention. The computer system of FIG. 1 is presented as an example to which this embodiment is applied. Referring to FIG. 1, a computer system may be provided with heterogeneous multi-processing (HMP) technology for real-time processing of high-performance application programs.

A computer system includes a processing unit including a graphic processing unit (GPU) 300 and a central processing unit (CPU) 400a and 400b for executing an application program. To support games with varying workload characteristics, it is necessary to manage the power of CPU clusters and GPUs together.

Conventional power management techniques derive the relationship between CPU utilization and FPS performance through off-line regression analysis, but in the case of heterogeneous multiprocessing computer systems, when both CPU clusters and GPUs are used, It is not easy to derive the relationship between CPU utilization and FPS performance through regression analysis.

Therefore, the present invention can perform dynamic voltage frequency scaling without performing a static regression analysis before execution of an application program, and efficiently manage the clock frequencies (operating frequencies) of a plurality of CPU clusters and the GPU to increase the power gain It presents a new technology that can be.

The dynamic frequency adjustment apparatus 100 adjusts the operating frequencies of the CPUs and the GPU based on the utilization of the CPUs and the GPU and the FPS information. To this end, the dynamic frequency adjustment apparatus 100 is configured to periodically obtain the usage rate and the FPS information in real time.

In one embodiment, the dynamic frequency adjustment apparatus 100 may receive a performance indicator (FPS) in real time from the information providing unit 200. The information providing unit 200 can be provided in the SurfaceFlinger process of the Android. In addition, the dynamic frequency adjustment apparatus 100 can receive the GPU usage rate and the CPU usage rate from the GPU driver 310 and the CPU driver 410 in real time.

Power management of the CPUs 400a and 400b and the GPU 300 can be independently performed by the two CPU management units 420a and 420b and one GPU management unit 320. [ The dynamic frequency adjustment apparatus 100 includes a CPU driver 410, CPU management units 420a and 420b, and a GPU 300 that are present in a kernel memory space for power management of the CPUs 400a and 400b and the GPU 300. [ Driver 310 and the GPU management unit 320 via the sysfs interface.

The dynamic frequency adjustment apparatus 100 determines the operating frequency of each processing unit (CPU, GPU) during the execution of the application program based on the utilization rate of the processing units (CPU, GPU) and the real time FPS fluctuation. 2 is a configuration diagram of a dynamic frequency adjusting apparatus according to an embodiment of the present invention. 1 and 2, the dynamic frequency adjusting apparatus 100 may include a learning unit 120, a tracking unit 140, and an identifying unit 160. [

The learning unit 120 calculates a minimum frequency that satisfies a maximum performance index (e.g., maximum achievable FPS) of the application program among the frequencies applicable to the processing unit (CPU, GPU) as a target frequency, Set upper and lower limits of target performance index range based on performance index.

For example, when the first CPU 400a of the processing units (CPU, GPU) is provided to the A7 CPU, the frequencies applicable to the first CPU 400a are 1000, 1100, 1200, 1300, and 1400 MHz. As another example, when the second CPU 400b is provided to the A15 CPU, the frequencies applicable to the second CPU 400b are 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900 and 2000 MHz. When the GPU 300 is provided as a MALI T628 GPU, the frequencies applicable to the GPU 300 are 177, 266, 350, 420, 480 and 543 MHz.

For example, when the first CPU 400a is used, a maximum FPS can be obtained at an operating frequency of 1200, 1300, or 1400 MHz. When the FPS is reduced below the maximum FPS at a frequency of 1000 or 1100 MHz, Is 1200 MHz, and this frequency is calculated as a target frequency for the first CPU 400a.

In one embodiment, the target performance index range may be determined by applying an upper limit and a lower limit margin set in advance to the maximum performance index. For example, when the maximum FPS is 30 and the upper and lower margin ranges are ± 2.5, the target performance index range is 27.5 to 32.5 (30 ± 2.5), and the upper limit value of the target performance index range is 32.5 and the lower limit value is 27.5 do.

The tracking unit 140 monitors whether the FPS measured in real time is out of the target performance index range and increases the operating frequency of the CPU or GPU if it is determined that the FPS is less than the lower limit of the target performance index range And requests the learning unit 120 to reset the target performance index range when it is determined that the target performance index range is larger than the upper limit value of the target performance index range. In addition, the tracking unit 140 may reduce the operating frequency so that the usage rate can be secured to a certain level or higher when the usage rate of the processing unit (CPU, GPU) is less than the threshold while the FPS belongs to the target FPS range.

The confirmation unit 160 compares the current performance indicator (current FPS) calculated for the application program after the operation frequency of the processing unit (CPU, GPU) is increased by the tracking unit 140 with the target performance indicator range and the previous FPS To determine whether the cause of the decrease in FPS is due to a gradual change in the workload or a sudden change in the application program. The specific operation and function of the learning unit 120, the tracking unit 140, and the confirmation unit 160 will be described later in detail with reference to FIG. 4 to FIG.

3 is an overall flowchart of a dynamic frequency adjustment method according to an embodiment of the present invention. Referring to FIG. 3, when an application program such as a 3D mobile game is executed, a DVFS (Dynamic Voltage Frequency Scaling) policy of each CPU core and a GPU is initialized to a user memory space in an initialization step S20, And the frequency table is created by reading the DVFS table of each core including the supported frequency values. After the initialization step S20, an update step S40 for reading information for dynamic frequency adjustment, for example, utilization of the CPUs and GPUs and FPS information is performed.

Next, a loading period of the application program is sensed (S60). Applications such as mobile games include loading status. The loading state can be detected by checking whether the FPS value has changed. The FPS value changes each time the SurfaceFlinger process updates information, but in the loading state, the frame is not updated for a while, and the SurfaceFlinger process is not called.

Therefore, if the FPS value does not change for a predetermined time, it can be determined that the FPS value is in the loading state. In the loading state, the CPU loads the game and the GPU becomes idle. Thus, if a loading state is detected, for fast loading, the clock frequency of the CPU may be set to a maximum value and the clock frequency of the GPU may be set to the minimum value (S80).

If loading is not detected in step S60, a dynamic frequency adjustment step (SlOO) is performed, and the clock frequencies of the CPUs and the GPU are adaptively determined according to the usage rate and the FPS. After the clock frequency is updated (S180), the above processes (S40 to S180) are repeated each time the set time (for example, 100 ms) elapses.

4 is a state diagram of step S100 of FIG. The state diagram of the dynamic frequency adjustment method according to the present embodiment includes three states: a learning state S120, a tracking state S140, and a checking state S160.

In order to calculate the initial target FPS range and the target frequency of the processing units (CPU, GPU), the learning state S120 is first entered. In the learning state (S120), the target FPS range is determined based on the maximum FPS that can be obtained. The maximum FPS at the beginning of application execution can be obtained when the maximum frequency is set in all the operating units.

The main purpose of the learning state (S120) is to minimize the power consumption by finding a new target FPS and setting the target frequencies to the minimum frequencies of the processing units satisfying the target FPS. When the target frequency is determined for the target FPS range and all the processing units, the learning state (S120) is switched to the tracking state (S140).

In the tracking state (S140), the clock frequency is adaptively adjusted so as to withstand short-term variations in scene complexity, while keeping the FPS performance within the target FPS range. To this end, it is determined whether the FPS calculated in real time satisfies the target FPS range.

If the FPS performance is out of the target FPS range in the tracking state (S140), the FPS violation is detected. If the measured FPS is less than the lower limit of the target FPS range in the tracking state (S140), it switches to the confirmation state (S160) to determine the cause of FPS reduction. In the tracking state S140, if the FPS is larger than the upper limit value of the target FPS range, the process returns to the learning state S120 for redetermining the target FPS range.

In the confirmation state (S160), the cause of FPS fluctuation (decrease) is investigated. If the FPS reduction is due to sudden scene changes, such as running another game or moving to a new scene, the current target FPS no longer guarantees the best performance, so the learning state (S120).

Alternatively, if it is determined that the cause of the FPS reduction is due to the gradual workload fluctuation of the processing unit (CPU, GPU), the FPS request can be adjusted to the current target FPS range only by increasing the frequency, ) To the tracking state (S140).

An example of calculating the performance index based on the FPS value calculated during execution of the application program in the learning state (S120), the tracking state (S140), and the confirmation state (S160) will be described first. Since the FPS value changes dynamically depending on the scene complexity and the system state of the frame, in order to secure the excellent performance, the FPS average value is calculated at regular intervals based on the history of the FPS value change during the execution of the application program, It is preferable to use it as an index for dynamic frequency control.

The FPS may be, for example, a SurfaceFlinger Can be measured by taking the inverse of the time interval between two calls of the process. However, if the FPS is measured at every frame update time, the FPS will fluctuate extensively for various reasons such as scene changes and irregular call delays in the SurfaceFlinger process.

To get a reliable and reliable FPS, SurfaceFlinger May be provided to update the FPS information intermittently without updating the FPS information at every call. After the FPS is updated, SurfaceFlinger The process grasps the number of frame updates for a predetermined time (e.g., 100 ms) and waits until a certain time has elapsed since the last update.

There is a trade-off relationship between the stability and the speed of adaptation according to the update period of the FPS. If the update interval is too short, adaptation will be difficult due to frequent fluctuations in FPS. On the other hand, if the update interval is long, the adaptation is delayed excessively and the fast scene complexity can not follow the change. The FPS can be calculated, for example, according to Equation 1 below.

[Equation 1]

In Equation 1, t _now is the SurfaceFlinger The current call time of the process t _prev is the previous call time of the SurfaceFlinger process, N _now is the SurfaceFlinger between the current call time and the previous call time The number of calls in the process (number of frame updates). SurfaceFlinger process can save the file in the calculated fps _ info FPS _e sysfs virtual file in Linux (Linux). Dynamic frequency control device reads fps _ info file, you can check the latest updates FPS _e information.

SurfaceFlinger There is a time delay in updating the clock frequency after reading the last FPS _e value recorded by the process, and the FPS _e value has also changed in the meantime. Therefore, when reading the FPS _e value from the fps_info file, SurfaceFlinger The previous FPS _e value recorded by the process may be read. In addition, the effect of the frequency adjustment appears after the delay time, and the dynamic frequency adjustment device and SurfaceFlinger Processes operate independently without synchronization, so the time delay changes dynamically. To solve this problem, after collecting five FPS _e samples as an example, the average FPS _avg value can be calculated according to the following Equation 2 by averaging the three FPS _e values excluding the maximum value and the minimum value.

[Equation 2]

In Equation 2, the FPS _sum is the _sum of 5 FPS _e samples, and FPS _max and FPS _min are the maximum FPS _e and the minimum FPS _e value, respectively, of 5 FPS _e samples.

5 is a time diagram showing a frame generation time point, an FPS update time point, and a frequency adjustment time point. Referring to FIG. 5, the FPS _e value is updated every set time (for example, 100 ms), and the frequency can be adjusted every set period (for example, 500 ms) in which five FPS _e values are calculated. In Figure 5, the initial FPS _e value after the frequency adjustment will be less than the next FPS _e value and will not be included in the average FPS _avg calculation.

FIG. 6 is a state diagram showing the learning state (S120) shown in FIG. 4 in detail. 2, 4 and 6, the internal state diagram of the learning state S120 includes four states of an initial state S122, a GPU state S124, a CPU state S126, and a decision state S128 .

In the initial state S122, the learning unit 120 determines a priority order for a plurality of processing units (CPU, GPU) based on the usage rates of a plurality of processing units (CPU, GPU). The high utilization rate of processing has a high priority to reduce the clock frequency.

When the priority order is determined for the processing units (CPU, GPU), the target frequencies can be calculated by sequentially determining the combination of the minimum frequencies for the processing units according to the priority order. The state moves from the initial state S122 to the GPU state S124 or the CPU state S126 depending on the usage rates of the processing units.

If the GPU has a higher priority than the CPU, the GPU state (S124) is first switched to the GPU state (S124) as shown by the solid line arrow in Fig. 6 to determine the minimum frequency for the GPU to calculate the target frequency, The CPU state (S126) is set to determine the minimum frequency for the CPU, and the target frequency is calculated.

Conversely, when the CPU has a higher priority than the GPU (when the CPU usage rate is higher than the GPU usage rate), the CPU state is first switched to the CPU state S126 as shown by the dotted arrow in Fig. 6 to determine the minimum frequency for the CPU The target frequency is calculated, the target FPS range is set, and then the GPU state (S124) is determined to determine the minimum frequency for the GPU to calculate the target frequency.

A method of calculating the target FPS range and the target frequency of the GPU and the CPU in the GPU state (S124), the CPU state (S126) will be described in more detail. At the beginning of application execution, the maximum FPS that can be obtained is calculated with the clock frequencies of all processors (CPU, GPU) set to the maximum value. At this time, FPS _{avg, which} is the average value of the FPS values measured within the set period, can be set to the maximum FPS.

In order to compensate for variations _avg FPS, the allowable range for the FPS _avg change is applied a target FPS range can be set. Then, in order to minimize power consumption, the FPS _avg is the extent kept in the range, reducing the clock frequency of the CPU and the GPU and, determining a minimum frequency at which the FPS _avg range maintained at the target frequency.

Once the target frequency range for the CPU and GPU is determined, it is switched to the determined state (S128). In the learning process, since the workload or the scene complexity of the game may change, it is confirmed in the determination state S128 that the current FPS still satisfies the target FPS range, and if it satisfies the operation of the processing units (CPU, GPU) It is switched to the tracking state (S140) in order to adjust the frequency. If the current FPS state does not satisfy the target FPS range in the determination state S128, the process moves to the initial state S122 of the learning state S120 and starts learning again.

7 is a flowchart of a tracking step constituting a dynamic frequency adjustment method according to an embodiment of the present invention. Referring to FIG. 2 and FIG. 7, in the tracking state (S140), when an FPS violation is detected, the target frequencies of the processing units are scaled to maintain the target FPS range.

In one embodiment, the tracking unit 140 determines whether the current FPS _avg is smaller than the lower limit value " min (target FPS) " of the target FPS range (S141, S142). If the current FPS _avg is smaller than the lower limit value of the target FPS range, the tracking unit 140 increases the operating frequency based on the current FPS _avg , the maximum FPS and the target frequency (S143), stores the current FPS _avg value and the clock frequencies The FPS reduction cause analysis unit 160 requests the confirmation unit 160 to analyze the cause of the FPS reduction. Thereafter, the cause of the FPS violation is checked in the confirmation state (S160).

In one embodiment, if the current FPS (current FPS _avg ) is smaller than the lower limit value " min (targetFPS) " of the target FPS range, the target frequencies Freq ' _c and Freq' _g of the CPU and GPU are expressed by Equation 3 and Equation 4 below. And can be scaled up accordingly.

[Equation 3]

[Equation 4]

- Freq ' _c , Freq' _g : Target frequency of CPU and GPU

- Freq _c , Freq _g : Current frequency of CPU and GPU

- FPS _target , FPS _current : Target FPS and current FPS

Since the frequencies allowed for each processing unit are limited, the target frequency typically does not match one of the allowable frequencies. At this time, the lowest frequency among the allowable frequencies above the target frequency is determined as the target frequency. After the clock frequency of the CPU, GPU is scaled, FPS _avg can be compared with the target FPS again. If the clock frequency of the CPU or GPU still needs to be changed, the frequency is changed stepwise adaptively.

If the FPS (present FPS _avg ) is larger than the upper limit value " max (targetFPS) " of the target FPS range (S144), it means that the FPS larger than the current target FPS can be obtained. After setting the maximum frequency as the operation frequency (S145), the learning unit 120 is requested to reset the target FPS range.

Even if the current FPS _avg is within the target FPS range, the current frequencies for the CPU or GPU may not be minimal due to variations in the workload of the game. In order to confirm this possibility, the usage rate of the processing units is monitored in real time in the tracking state (S140). The utilization rate is an index for identifying the performance bottleneck of the processing apparatus.

If the performance index FPS satisfies the target performance index range (S142, S144), the usage rate of the processing device is compared with the threshold value MIN_THRESHOLD. If the usage rate is smaller than the threshold value MIN_THRESHOLD, (S147).

In one embodiment, the GPU utilization value is provided by the GPU driver, and the sysfs The utilization in the file system You can read from the file. Since CPU utilization information is not provided by the system, cpufreq Set the kernel driver to calculate the CPU core usage rate. The core usage rate can be calculated, for example, by dividing the total execution time of each core by the required time. Similar to the FPS information, usage rate information may also be updated periodically (e.g., 100 ms period).

For example, in the case of a Cortex- A7 cluster, the usage rates can be calculated by dividing the usage rates of the four cores by a set value corresponding to the background workload of the A7 cluster (for example, 50). On the other hand, for Cortex-A15 cores cores, since core migration can occur during the utilization ratio measurement period, the utilization ratio can be calculated by adding the two largest utilization values.

You can compare the utilization of the CPU and GPU with the threshold value to see if the stage of the current game is concentrated on the CPU or GPU, and the clock frequency is adjusted based on the usage rate. If the usage rate of the processing unit is low, it is assumed that the core is not a performance bottleneck, and the utilization rate of the processing unit is lowered in a balanced manner with other processing units (S147).

The threshold value may be derived from a threshold value table used by the GPU management unit. This embodiment helps to ensure that the processing unit is not kept at a clock frequency that is too high above the required performance when a scene change occurs. If this scaling down leads to FPS loss, the operating frequency of the processing unit can be recovered immediately to the next DVFS period (e.g., 500 ms), so that it does not significantly affect the service performance quality.

FIG. 8 is a flow chart of an identifying step for configuring a dynamic frequency adjusting method according to an embodiment of the present invention. Referring to FIG. 2 and FIG. 8, the main function of the confirmation state S160 is to confirm whether the FPS can be recovered within the target FPS range by DVFS scaling performed in the previous tracking state.

In one embodiment, the verification unit 160 determines whether the current performance indicator (FPS _avg value after frequency scaling) satisfies the target performance indicator range. If it is determined that the current performance indicator does not satisfy the target performance indicator range, And compares it with the calculated past performance index (S161, S162).

If the current performance index is lower than the past performance index (S163), the main reason for the decrease in FPS is a performance bottleneck due to an increase in the workload of the processing apparatus. Therefore, the operating frequency is adjusted based on the current performance index (S164) and request the learning unit 120 to reset the target performance index range. If the CPU or GPU is the bottleneck, increasing the frequency will increase FPS performance. Then, the state is changed back to the tracking state (S140), and as the frequency further increases, the FPS will also increase.

Alternatively, if such scaling is not sufficient to recover the FPS, it is assumed that the gain workload has changed abruptly. In this case, it is necessary to set a new target FPS to cope with unexpected fluctuations. At this time, before going to the learning state (S120), the frequency is again scaled. If learning is performed starting from the scaled frequency values rather than the maximum frequencies, learning time required for finding the target performance index range and target frequency in the learning state (S120) can be reduced. To avoid increasing the frequency of the processing unit being idle, the frequency scaling in the confirmation state (S160) may be set not to apply to processing units whose utilization rate is less than a reference value (e.g., 50%).

9 to 12 are graphs showing performance (power gain, FPS) of the dynamic frequency adjustment method according to the embodiment of the present invention. The software and hardware specifications used to verify the performance of the present invention are shown in Table 1 below.

FIG. 9 is a graph showing power and FPS (proposed W, proposed FPS) of the present invention, power of Comparative Example 1 and default power (FPS) of six application programs (Edge of tomorrow, Asphalt 8, Nova, Epic, Modern, Implosion) FIG. 10 and FIG. 11 are graphs showing power and FPS variation of the present invention and Comparative Example 1 for "Edge of tomorrow" and "Nova" games, respectively. Comparative Example 1 is a result of executing an application program with the management section of the A7 cluster and the A15 cluster being set as default.

As shown in FIGS. 9 to 11, in the case of the dynamic frequency adjustment method according to the embodiment of the present invention, the FPS performance is equivalent to that of six application programs, and four out of six applications (Edge of tommorrow , Nova, Modern, Implosion), and power consumption was comparable to that of Comparative Example 1 for two application programs (Asphalt8, Epic).

FIG. 12 shows the power and FPS (proposed W, proposed FPS) of the present invention and the power and FPS (predictive W, predictive FPS) of Comparative Example 2 for four application programs (Edge of tomorrow, Nova, Epic, Respectively. Comparative Example 2 is described in <A. Pathania, AE Irimiea, A. Prakash, and T. Mitra, "Power-Performance Modeling of Mobile Gaming Workloads on Heterogeneous MPSoCs", in Proceedings of the 52Nd Annual Design Automation Conference , DAC '15, (New York, NY, USA), pp. 201: 1-201: 6, ACM, 2015. &gt;. As shown in FIG. 12, the dynamic frequency control method according to the embodiment of the present invention exhibits power consumption characteristics equivalent to or lower than that of the comparative example 2 while maintaining the FPS performance.

The method according to an embodiment of the present invention can be realized in a general-purpose digital computer that can be created as a program that can be executed by a computer and operates the program using a computer-readable recording medium. The computer readable recording medium may be a volatile memory such as SRAM (Static RAM), DRAM (Dynamic RAM), SDRAM (Synchronous DRAM), ROM (Read Only Memory), PROM (Programmable ROM), EPROM (Electrically Programmable ROM) Non-volatile memory such as EEPROM (Electrically Erasable and Programmable ROM), flash memory device, Phase-change RAM (PRAM), Magnetic RAM (MRAM), Resistive RAM (RRAM), Ferroelectric RAM But are not limited to, optical storage media such as CD ROMs, DVDs, and the like.

It is to be understood that the above-described embodiments are provided to facilitate understanding of the present invention, and do not limit the scope of the present invention, and it is to be understood that various modified embodiments are also within the scope of the present invention. It is to be understood that the technical scope of the present invention should be determined by the technical idea of the claims and the technical scope of protection of the present invention is not limited to the literary description of the claims, To the invention of the invention.

100: Dynamic frequency adjusting device
120:
140:
160:
200: Information provision
300: GPU
400a, 400b: CPU

Claims (11)

Determining a minimum frequency that meets a maximum performance index of the application program among allowable frequencies for a processing apparatus that executes the application program, calculating a target frequency for the processing apparatus, and calculating, based on the maximum performance index, And adjusts the operating frequency of the processing apparatus based on the performance index calculated during execution of the application program, the target performance index range, and the target frequency. The method according to claim 1,
The dynamic frequency adjusting device comprises:
A learning unit that calculates a minimum frequency that satisfies the maximum performance index among the allowable frequencies for the processing apparatus at the target frequency and determines an upper limit value and a lower limit value of the target performance index range based on the maximum performance index;
A tracking unit for monitoring whether the performance index calculated during execution of the application program is out of the target performance index range and increasing the operation frequency of the processing unit or requesting the learning unit to re-determine the target performance index range; And
And a confirmation unit for comparing the current performance index calculated for the application program with the target performance indicator range after the operating frequency of the processing unit is increased by the tracking unit and analyzing the cause of the reduction of the performance indicator Frequency adjusting device. 3. The method of claim 2,
Wherein the processing apparatus includes a plurality of processing sections,
Wherein,
Determining a priority order for the plurality of processors based on a usage rate of the plurality of processors, determining a combination of minimum frequencies satisfying the maximum performance index for the plurality of processors sequentially according to the priority, / RTI &gt; The method according to claim 2 or 3,
The tracking unit includes:
If the performance index is smaller than the lower limit value of the target performance index range, the operating frequency is increased based on the performance index, the maximum performance index and the target frequency, Request,
And sets the maximum frequency among the frequencies to the operating frequency when the performance index is larger than the upper limit value of the target performance index range and requests the learning unit to reset the target performance index range and the target frequency,
And compares the utilization rate of the processing apparatus with a preset threshold value when the performance index satisfies the target performance index range and decreases the operation frequency when the utilization rate is smaller than the threshold value. The method according to claim 2 or 3,
The checking unit,
Comparing the current performance indicator with a past performance indicator calculated before the operating frequency is increased when it is determined that the current performance indicator does not satisfy the target performance indicator range,
A dynamic frequency adjustment unit for adjusting the operating frequency based on the current performance index and requesting the learning unit to reset the target performance index range when it is determined that the current performance index is lower than the past performance index; . Determining a minimum frequency that meets the maximum performance index of the application program among the allowable frequencies in the processing apparatus that executes the application program to calculate a target frequency for the processing apparatus, and based on the maximum performance index, ; And
And a tracking step of adjusting an operating frequency of the processing apparatus based on the performance index calculated during execution of the application program, the target performance index range, and the target frequency. The method according to claim 6,
And analyzing the cause of the decrease in the performance index by comparing the current performance index calculated for the application program with the target performance index range after the operating frequency of the processing apparatus is increased in the tracking step, Frequency adjustment method. 8. The method of claim 7,
Wherein the tracking step comprises:
And when the performance index is smaller than the lower limit value of the target performance index range, the operating frequency is increased based on the performance index, the maximum performance index, and the target frequency, ;
Setting the maximum frequency among the frequencies to the operating frequency of the processing device when the performance index is larger than the upper limit value of the target performance index range, and then returning the target performance index range and the target frequency to the learning step Switching; And
And comparing the usage rate of the processing device with a preset threshold value when the performance index satisfies the target performance index range and decreasing an operation frequency of the processing device when the utilization rate is smaller than the threshold value / RTI &gt; 8. The method of claim 7,
Wherein,
Comparing the current performance indicator with the target performance indicator range, and when the current performance indicator is determined not to satisfy the target performance indicator range, comparing the current performance indicator with a past performance indicator calculated before the operation frequency is increased ; And
Wherein the control unit adjusts the operating frequency based on the current performance index, the maximum performance index, and the target frequency, and then, if the current performance index is determined to be smaller than the past performance index, And switching to the learning step. The method according to claim 6,
In the learning step,
Determining a priority order for the plurality of processing units based on utilization rates of the plurality of processing units constituting the processing apparatus; And
And calculating target frequencies by sequentially determining a combination of minimum frequencies satisfying the maximum performance index for the plurality of processors in accordance with the priority order. A computer-readable recording medium on which a program for executing the dynamic frequency adjustment method according to any one of claims 6 to 10 is recorded.

Images