WO2015165045A1 - Power control method and apparatus - Google Patents

Abstract

A power control method and apparatus, which can ensure performance of an HPE group. The method comprises: sampling, at least once, a performance indicator of each HPE in a to-be-controlled HPE group comprising at least two heterogeneous processing engines (HPEs) comprised in a heterogeneous processor, and a working frequency of the HPE corresponding to the performance indicator; determining a performance characterization value of the to-be-controlled HPE group corresponding to each sampling according to the performance indicator value of each HPE obtained in each sampling; determining, according to a working frequency value of each HPE obtained in each sampling and the performance characterization value of the to-be-controlled HPE group corresponding to each sampling, a correlation coefficient between the working frequency value of each HPE and the performance characterization value of the to-be-controlled HPE group, wherein the larger the correlation coefficient, the greater the effect of the working frequency value characterizing the corresponding HPE on the performance characterization value of the to-be-controlled HPE group; and performing power control on each HPE according to the determined correlation coefficient.

Description

 Power control method and device

Technical field

 The present invention relates to the field of computer technologies, and in particular, to a power control method and apparatus for a heterogeneous processing engine in a heterogeneous processor. Background technique

 With the development of technology, heterogeneous processors have been widely used. The schematic diagram is shown in Figure 1. A heterogeneous processor will have multiple HPEs (Heterogeneous Processing Engines). For example, HPE1 is a CPU. (Central Processing Unit, CPU) Corresponding HPE, HPE2 is HPE corresponding to GPU (Graphic Processing Unit), HPE1 and HPE2 are different types of HPE. Multiple HPEs in a heterogeneous processor can form an application for an HPE group. In an HPE group, one HPE acts as the main processor, and other HPEs act as an acceleration processing unit to share the memory, heat dissipation, and power of the heterogeneous processor. HPE Each HPE in the group has an impact on the performance of the application running on the HPE group.

 In the prior art, when power control is performed on the HPE in the heterogeneous processor, the BAPM (Bidirectional Application Power Management) algorithm is usually used to allocate power to each of the heterogeneous processors according to the heat space margin. However, existing power control schemes tend to maximize the heat dissipation capabilities of heterogeneous processors and do not guarantee the performance of the group. Summary of the invention

 Embodiments of the present invention provide a power control method and apparatus for a 异构 group in a heterogeneous processor, which can ensure the performance of the ΗΡΕ group.

 In a first aspect, a power control method is provided, including:

Performing at least one sampling of a performance indicator of each of the to-be-controlled groups of the heterogeneous processor including at least two heterogeneous processing engines, and an operating frequency of the pair corresponding to the performance indicator; Determining the performance characterization value of the HPE group to be controlled corresponding to each sample according to the performance index value of each HPE obtained by each sampling;

 Determine the operating frequency value of each HPE obtained for each sample and the performance characterization value of the HPE group to be controlled corresponding to each sample, and determine the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled. Correlation coefficient; wherein, the greater the correlation coefficient, the greater the effect of characterizing the operating frequency value of the corresponding HPE to control the performance characterization value of the HPE group;

 Power control is performed for each HPE based on the correlation between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 With reference to the first aspect, in a first possible implementation manner, determining a performance characterization value of the to-be-controlled HPE group corresponding to each sample according to the performance index value of each HPE obtained each time, including: The performance index value of each HPE sampled is determined, and the performance characterization value of each HPE corresponding to each sample is determined;

 Based on the performance characterization values of each HPE corresponding to each sample, the performance characterization values of the HPE group to be controlled corresponding to each sample were determined.

 With reference to the first possible implementation manner of the first aspect, in a second possible implementation manner, determining performance characterization of each HPE corresponding to each sample according to the performance index value of each HPE obtained by each sampling Values, including:

 For each HPE, when the performance index value of the HPE sampled is one, it is determined that the performance characterization value of the HPE corresponding to each sample is the performance index value of the HPE obtained by each sampling;

 When the performance index value of the sampled HPE is at least two, the analytic hierarchy process, the network analysis method or the multi-attribute performance decision method is used to calculate at least two performance index values of the HPE obtained by each sampling. , to determine the performance characterization value of the HPE corresponding to each sample.

 With reference to the first possible implementation manner of the first aspect, in a third possible implementation manner, determining performance characterization of the to-be-controlled HPE group corresponding to each sampling according to the performance characterization value of each HPE corresponding to each sampling Values, including:

Use analytic hierarchy process, network analysis method or multi-attribute effectiveness decision method to correspond to each sampling The performance characterization values of each of the HPEs are calculated to determine the performance characterization values of the HPE group to be controlled corresponding to each sample.

 With reference to the first aspect, the first possible implementation manner of the first aspect, the second possible implementation manner of the first aspect, or the third possible implementation manner of the first aspect, in the fourth possible implementation manner Determining, between each working frequency value of each HPE obtained by each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling, determining the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled Correlation coefficients, including:

 The multivariate linear regression method, the multivariate nonlinear regression method or the complex correlation coefficient method is used to calculate the operating frequency value of each HPE obtained for each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sample to determine A correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 With reference to the first aspect, the first possible implementation of the first aspect, the second possible implementation of the first aspect, the third possible implementation of the first aspect, or the fourth possible aspect of the first aspect In the fifth possible implementation manner, before performing power control on each HPE, the following includes:

 It is determined that the difference between the current memory bandwidth of the heterogeneous processor and the theoretical memory bandwidth is greater than the preset bandwidth difference.

 With reference to the first aspect, the first possible implementation of the first aspect, the second possible implementation of the first aspect, the third possible implementation of the first aspect, the fourth possible The implementation manner, or the fifth possible implementation manner of the first aspect, in the sixth possible implementation manner, according to a correlation coefficient between an operating frequency value of each HPE and a performance representative value of the HPE group to be controlled, Power control for each HPE, including:

 Determining, according to a preset relationship between the coefficient interval and the working frequency value, an operating frequency value corresponding to a coefficient interval in which a correlation coefficient between each HPE operating frequency value and a performance representative value of the HPE group to be controlled is located; wherein, the coefficient The larger the coefficient value in the interval, the larger the operating frequency value corresponding to the coefficient interval;

Power control is performed for each HPE based on the determined operating frequency value. In conjunction with the sixth possible implementation of the first aspect, in a seventh possible implementation, before performing power control on each of the HPEs according to the determined operating frequency value, the method further includes:

 After pre-judging the power control of each HPE based on the determined operating frequency value, the power of the heterogeneous processor does not exceed the maximum power allowed by the heterogeneous processor.

 In a second aspect, a power control apparatus is provided, including:

 a sampling unit, configured to perform performance indicators of each HPE in the HPE group to be controlled that includes at least two heterogeneous processing engines HPE in the heterogeneous processor, and the operating frequency of the HPE corresponding to the performance indicator at least once Sampling

 a first determining unit, configured to determine, according to a performance index value of each HPE obtained for each sample, a performance characterization value of the HPE group to be controlled corresponding to each sampling;

 a second determining unit, configured to determine an operating frequency value of each HPE and a performance of the HPE group to be controlled according to an operating frequency value of each HPE obtained by each sampling and a performance characterization value of the HPE group to be controlled corresponding to each sampling a correlation coefficient between the characterization values; wherein, the larger the correlation coefficient, the greater the effect of characterizing the operating frequency value of the corresponding HPE to control the performance characterization value of the HPE group;

 A control unit is configured to perform power control for each HPE based on a correlation coefficient between an operating frequency value of each HPE and a performance characterization value of the HPE group to be controlled.

 With reference to the second aspect, in a first possible implementation, the first determining unit is specifically configured to determine, according to performance index values of each type of HPE obtained by each sampling, performance of each HPE corresponding to each sampling. Characterizing the value; determining the performance characterization value of the HPE group to be controlled corresponding to each sample according to the performance characterization value of each HPE corresponding to each sampling.

With reference to the first possible implementation manner of the second aspect, in a second possible implementation manner, the first determining unit is specifically configured to: for each HPE, when the sampled the performance index value of the HPE is In one case, it is determined that the performance characterization value of the HPE corresponding to each sample is the performance index value of the HPE obtained by each sample; when the performance index value of the sampled HPE is at least two, The analytic hierarchy process, the network analysis method or the multi-attribute performance decision method is used to calculate at least two performance index values of the HPE obtained by each sampling to determine the performance characterization value of the HPE corresponding to each sample. With reference to the first possible implementation manner of the second aspect, in a third possible implementation, the first determining unit is specifically configured to use an analytic hierarchy process, a network analysis method, or a multi-attribute performance decision method. The performance characterization values of each HPE corresponding to the second sample are calculated to determine the performance characterization values of the HPE group to be controlled corresponding to each sample.

 With reference to the second aspect, the first possible implementation manner of the second aspect, the second possible implementation manner of the second aspect, or the third possible implementation manner of the second aspect, in the fourth possible implementation manner The second determining unit is specifically configured to adopt a multiple linear regression method, a multiple nonlinear regression method or a complex correlation coefficient method, and the working frequency value of each HPE obtained for each sampling and the corresponding control to be controlled for each sample The performance characterization values of the HPE group were calculated to determine the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 With reference to the second aspect, the first possible implementation of the second aspect, the second possible implementation of the second aspect, the third possible implementation of the second aspect, or the fourth possible aspect of the second aspect In a fifth possible implementation manner, the control unit is further configured to: before performing power control on each HPE, determine that a difference between a current memory bandwidth and a theoretical memory bandwidth of the heterogeneous processor is greater than The preset bandwidth difference.

 With reference to the second aspect, the first possible implementation of the second aspect, the second possible implementation of the second aspect, the third possible implementation of the second aspect, and the fourth possible aspect of the second aspect The implementation manner, or the fifth possible implementation manner of the second aspect, in the sixth possible implementation manner, the control unit is specifically configured to determine, according to the correspondence between the preset coefficient interval and the working frequency value, The working frequency value corresponding to the coefficient interval in which the correlation coefficient between the operating frequency value of the HPE and the performance representative value of the HPE group to be controlled is located; wherein, the larger the coefficient value in the coefficient interval, the more the operating frequency value corresponding to the coefficient interval Large; power control for each HPE based on the determined operating frequency value.

With reference to the sixth possible implementation manner of the second aspect, in a seventh possible implementation, the control unit is further configured to perform pre-determination according to the determined operating frequency value before performing power control on each HPE After the determined operating frequency value is power controlled for each HPE, the power of the heterogeneous processor does not exceed the maximum power allowed by the heterogeneous processor. According to the power control method provided by the first aspect, the power control apparatus provided by the second aspect, the performance index of each HPE in the HPE group to be controlled, and the operating frequency of the HPE corresponding to the performance index are obtained according to the The sample value determines the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled. The larger the correlation coefficient, the role of the operating frequency value corresponding to the HPE to control the performance characterization value of the HPE group. The larger, that is, the greater the effect of the operating frequency of the HPE to control the performance of the HPE group. Therefore, the power is applied to each HPE according to the correlation coefficient between the operating frequency value of each HPE and the performance representative value of the HPE group to be controlled. Control can guarantee the performance of the HPE group to be controlled. DRAWINGS

 The drawings are intended to provide a further understanding of the invention, and are intended to be a part of the description of the invention. In the drawings: Figure 1 is a schematic structural diagram of a heterogeneous processor;

 2 is a flowchart of a power control method according to an embodiment of the present invention;

 3 is a detailed flowchart of a power control method according to Embodiment 1 of the present invention;

 4 is a detailed flowchart of a power control method according to Embodiment 2 of the present invention;

 FIG. 5 is a schematic structural diagram of a power control apparatus according to Embodiment 3 of the present invention; FIG.

 FIG. 6 is a schematic structural diagram of a power control device according to Embodiment 4 of the present invention. detailed description

 In order to provide a power control method and a device for ensuring the performance of the HPE group in the heterogeneous processor, the present invention provides a power control method and apparatus. The following describes the preferred embodiment of the present invention with reference to the accompanying drawings. The preferred embodiments described herein are for illustrative purposes only and are not intended to limit the invention. And in the case of no conflict, the features in the embodiments and the embodiments of the present application can be combined with each other.

An embodiment of the present invention provides a power control method. As shown in FIG. 2, the method includes the following steps: Step 201: To-be-controlled HPE including at least two heterogeneous processing engines HPE in a heterogeneous processor The performance index of each HPE in the group and the working frequency of the HPE corresponding to the performance indicator are sampled at least once;

 Step 202: Determine, according to the performance index value of each HPE obtained in each sample, a performance characterization value of the HPE group to be controlled corresponding to each sample;

 Step 203: Determine, according to an operating frequency value of each HPE obtained by each sampling, and a performance characterization value of the HPE group to be controlled corresponding to each sampling, determine an operating frequency value of each HPE and a performance characterization value of the HPE group to be controlled. Correlation coefficient; wherein, the larger the correlation coefficient is, the greater the effect of characterizing the operating frequency value of the corresponding HPE to control the performance characterization value of the HPE group;

 Step 204: Perform power control on each HPE according to a correlation coefficient between an operating frequency value of each HPE and a performance characterization value of the HPE group to be controlled.

 The correlation coefficient between the operating frequency value of the HPE and the performance characterization value of the HPE group to be controlled characterizes the effect of the operating frequency value of the corresponding HPE to control the performance characterization value of the HPE group, that is, the working frequency of the corresponding HPE is characterized. The size of the effect of controlling the performance of the HPE group. The larger the correlation coefficient, the greater the effect of the HPE's operating frequency on the performance of the HPE group; conversely, the smaller the correlation, the smaller the effect of the HPE's operating frequency on the performance of the HPE group.

 It can be seen that, by using the power control method provided by the embodiment of the present invention, power control is performed on the HPE according to a correlation coefficient between an operating frequency value of each HPE and a performance representative value of the HPE group to be controlled, compared to the power of the prior art. The control scheme can ensure the performance of the HPE group to be controlled, and thus the application performance of the HPE group to be controlled can be guaranteed, and the user experience is better.

 The power control scheme provided by the embodiment of the present invention will be described in detail below with reference to the accompanying drawings.

 Example 1

 FIG. 3 is a flowchart of a method for controlling power of an HPE in a heterogeneous processor according to Embodiment 1 of the present invention, which specifically includes:

 Step 301: For each HPE in the HPE group to be controlled, the performance index of the HPE and the working frequency of the HPE corresponding to the performance index are sampled.

Wherein, the number of sampling times may be one time, preferably, the number of sampling times is multiple times, by expanding the data The sample makes the power control more reasonable.

 In actual implementation, a tool similar to CodeXL can be used to periodically collect the value of the relevant performance counter corresponding to each HPE and the corresponding operating frequency value, and use the value of the performance counter as the performance index value of the corresponding HPE. An HPE may have one operating frequency or multiple operating frequencies. If an HPE has multiple operating frequencies, the average operating frequency can be taken as the working frequency of the HPE.

 Among them, an HPE may contain only one performance indicator, and may also include multiple performance indicators. For example, the HPE performance indicators corresponding to the GPU mainly include the following seven types:

 ALU Insts PTI: Compute instructions per thousand instructions, the number of calculation instructions per thousand instructions in the arithmetic unit (ALU);

 ALU Busy: The percentage of GPU Time ALU instructions are processed, the percentage of time in the image processing unit (GPU) that processes the arithmetic logic unit instructions;

 ALU Fetch Ratio: The ratio of ALU prefetch instructions;

 GPU Clock Busy: GPU utilization;

 Write Unit Stalled: The % of GPU Time main memory write/store unit is stalled, the main memory write/storage unit is suspended as a percentage of the GPU time;

 Fetch Unit Stalled: The % of GPU Time main memory fetch/load unit is stalled , the main memory acquisition / import unit is suspended as a percentage of the GPU time;

 Fetch from memory PTI: Main memory reads per thousand instructions, the number of main memory reads per thousand instructions.

 For HPEs with multiple performance metrics, one, more, or all of the performance metrics can be sampled based on actual conditions.

 That is, the performance index value of the HPE obtained in the step 301 may be one type or two or more types.

In the embodiment of the present invention, assuming that the HPE group to be controlled includes m kinds of HPEs and performs n times of sampling, the operating frequency values of each type of HPE obtained by each sampling are as follows:

 Wherein, the working frequency value obtained by sampling the operating frequency of the jth HPE at the i-th time, 1 < i < n, 1 j m.

 Assuming that each HPE is smeared on the s performance metrics, the performance characterization values for each HPE obtained for each sample are shown in the following table:

Where r _ljk represents the performance index value obtained by the i-th time for the kth performance index of the jth HPE, lks.

 Step 302: Determine, according to the performance index value of each HPE obtained by each sampling, a performance characterization value of each HPE corresponding to each sampling.

For each HPE, when the performance index value of the HPE obtained by the sample is one, it can directly determine that the performance characterization value of the HPE corresponding to each sample is the performance index value of the HPE obtained by each sampling. When the performance index value of the HPE obtained by the sample is at least two, the analytic hierarchy process, the network analysis method, or the multi-attribute performance decision method may be used to obtain at least two performance indexes of the HPE obtained for each sampling. The value is calculated to determine the performance table of the HPE corresponding to each sample. Value.

 That is, the performance characterization values of each HPE obtained by each sampling in the above table 301 can be determined as follows:

Where R _3⁄4 represents the performance characterization value of the jth type of HPE corresponding to the ith sampling.

 Step 303: Determine, according to a performance characterization value of each HPE corresponding to each sampling, a performance characterization value of the HPE group to be controlled corresponding to each sampling, and specifically adopt an analytic hierarchy process, a network analysis method, or a multi-attribute performance decision method. The performance characterization values of each HPE corresponding to each sample are calculated to determine the performance characterization value of the HPE group to be controlled corresponding to each sample.

 The above-mentioned analytic hierarchy process, network analysis method or multi-attribute performance decision-making method is merely an example, and is a commonly used decision-making method, and is not intended to limit the present invention. Other decision methods in the prior art may also be used as an implementation algorithm for determining the performance characterization value of the HPE in step 302 and determining the performance characterization value of the HPE group to be controlled in step 303.

That is, by performing the performance characterization values of each HPE corresponding to each sample determined in the above table 302, it can be determined that the performance characterization values of the HPE group to be controlled corresponding to each sample are as follows: The number of sampling times to be controlled by the HPE group performance characterization value

 First sampling Pi

 2nd sampling P2

The first undersampling n where _P1 represents the performance characterization value of the HPE group to be controlled corresponding to the ith sampling.

 Step 304: Determine, according to an operating frequency value of each HPE obtained by each sampling, and a performance characterization value of the HPE group to be controlled corresponding to each sampling, determine an operating frequency value of each HPE and a performance characterization value of the HPE group to be controlled. The correlation coefficient may be multivariate linear regression method, multiple nonlinear regression method or complex correlation coefficient method, and the operating frequency value of each HPE obtained for each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling. A calculation is made to determine the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 The above multiple linear regression method, multivariate nonlinear regression method or complex correlation coefficient method is merely an example, and is a commonly used mathematical statistics method, and is not intended to limit the present invention. Other mathematical statistics methods in the prior art may also be used as an implementation algorithm for determining correlation coefficients in step 304.

That is, the operating frequency value of each HPE obtained by each sampling in the above table 301, and the performance characterization value of the HPE group to be controlled corresponding to each sampling determined in the above table 303, can determine the type of each HPE. The correlation coefficient between the operating frequency value and the performance characterization value of the HPE group to be controlled is shown in the following table:

 Where b'j represents the correlation coefficient between the operating frequency value of the jth HPE and the performance characterization value of the HPE group to be controlled.

 Step 305: Determine whether a difference between a current memory bandwidth of the heterogeneous processor and a theoretical memory bandwidth is greater than a preset bandwidth difference.

 When it is determined that the difference between the current memory bandwidth and the theoretical memory bandwidth of the heterogeneous processor is not greater than the preset bandwidth difference, the current memory bandwidth of the heterogeneous processor is close to the theoretical memory bandwidth, and then proceeds to step 306;

 When it is determined that the difference between the current memory bandwidth and the theoretical memory bandwidth of the heterogeneous processor is greater than the preset bandwidth difference, the current memory bandwidth of the heterogeneous processor is not close to the theoretical memory bandwidth. In this case, the process proceeds to step 307.

 The preset bandwidth difference can be set by the user according to actual application scenario requirements.

 Step 306: Control power reduction of at least one HPE, and the power control process ends.

 When it is determined that the difference between the current memory bandwidth of the heterogeneous processor and the theoretical memory bandwidth is not greater than the preset bandwidth difference, it indicates that the memory bandwidth of the heterogeneous processor has become a bottleneck. At this time, by reducing the power of the HPE, the memory is released. Bandwidth can increase the energy efficiency ratio of the system.

 In practice, HPE power can be reduced by reducing the operating frequency or operating voltage of the HPE.

 Further, in this step 306, the power reduction of part of the HPE can be controlled, and the power reduction of all the HPEs can also be controlled, which can be determined according to actual conditions.

Step 307: Determine each type of HPE according to a correspondence between a preset coefficient interval and an operating frequency value. The working frequency value corresponding to the coefficient interval in which the correlation coefficient between the operating frequency value and the performance representative value of the HPE group to be controlled is located; wherein, the larger the coefficient value in the coefficient interval, the larger the operating frequency value corresponding to the coefficient interval.

 In the embodiment 1 of the present invention, the coefficient interval may specifically include three:

 Coefficient interval 1. an interval smaller than the first specified coefficient value;

 Coefficient interval 2, an interval greater than or equal to the first specified coefficient value and less than the second specified coefficient value; a coefficient interval 3, an interval greater than or equal to the second specified coefficient value;

 Wherein, the second specified coefficient value is greater than the first specified coefficient value, that is, three coefficient intervals are divided by using two specified coefficient values.

 The preset coefficient interval 1 corresponds to the operating frequency value fl, the coefficient interval 2 corresponds to the operating frequency value of £2, and the coefficient interval 3 corresponds to the operating frequency value G, where ί1< £2< β.

 In other embodiments of the present invention, the coefficient interval may specifically include two, four or other numbers, that is, the coefficient interval may specifically include X, and X coefficient intervals are divided by X-1 designated coefficient values, wherein , X is a positive integer greater than or equal to 2.

 Step 308: Pre-determine whether the power of the heterogeneous processor meets the power requirement after performing power control on each of the chirps according to the determined operating frequency value, that is, after determining the power control of each chirp according to the determined operating frequency value, the heterogeneous Whether the power of the processor exceeds the maximum power allowed by the heterogeneous processor.

 After determining that the power of each of the ports is controlled according to the determined operating frequency value, the power of the heterogeneous processor meets the power requirement, and the power of the heterogeneous processor does not exceed the maximum power allowed by the heterogeneous processor, proceeding to step 309;

 When it is determined that the power of each heterogeneous power is controlled according to the determined operating frequency value, the power of the heterogeneous processor does not satisfy the power requirement, and when the power of the heterogeneous processor exceeds the maximum power allowed by the heterogeneous processor, the process proceeds to step 310.

Step 309: Perform power control on each of the ports according to the operating frequency value determined in step 307, and the power control process ends. In step 307, a working frequency value of a corresponding level is determined for each HPE. When the step 309 is implemented, the working frequency of each HPE can be directly controlled to a determined working frequency value, thereby achieving the purpose of power control, and also by controlling each The operating voltage of the HPE is such that the operating frequency of each HPE is a determined operating frequency value for power control purposes.

 Step 310: Reduce the operating frequency value of the at least one HPE determined in step 307, and perform power control on each HPE according to the reduced operating frequency value.

 When it is determined that power control of each HPE is performed according to the determined operating frequency value, if the power of the heterogeneous processor does not meet the power requirement, that is, power control of each HPE according to the determined operating frequency value may result in a heterogeneous processor. The power exceeds the maximum power allowed by the heterogeneous processor. At this time, the power control of each HPE cannot be performed according to the operating frequency value determined in step 307, but the operating frequency of some HPEs or all HPEs determined in step 307 should be lowered. The value is then power controlled for each HPE based on the reduced operating frequency value.

 It can be seen that the power control method provided by Embodiment 1 of the present invention considers the relationship between each HPE and the performance of each HPE to the performance of the HPE group when performing power control, and can ensure the performance of the HPE group, and also considers the heterogeneity. The memory bandwidth and power requirements of the processor can further improve the rationality and reliability of the power control. Bright.

 Example 2:

 In the second embodiment of the present invention, two HPEs are included in the HPE group to be controlled, and each HPE has two performance indexes.

 FIG. 4 is a flowchart of a power control method according to Embodiment 2 of the present invention, which specifically includes: Step 401: For each HPE, two performance indicators of the HPE and the HPE corresponding to the performance index The working frequency is sampled twice.

 That is, in the second embodiment of the present invention, the HPE type m=2, the number of sampling times n=2, and the sampling type of the performance index of each HPE is s=2.

Assume that the operating frequency values of each HPE obtained for each sample are as follows:

 The performance indicator values for each sample are shown in the following table:

 Step 402: Determine, according to the performance index value of each enthalpy obtained by each sampling, an analytic hierarchy process to determine a performance characterization value of each 对应 corresponding to each sampling, specifically:

Rii ⁼ Xi ^x rin+X2 ^x rii2; Ri2 ⁼ i' ^x ri2i+X2' ^x ri22;

X₂'、 Xj", X₂"、 ΧΓ、 X₂"'为权重系数， 权重系数 的具体确定方法为现有技术， 在此不再详述。 Where R _y represents the performance characterization value of the jth type of HPE corresponding to the ith sampling, and r _yk represents the performance index value obtained by the first time for the kth performance index of the jth HPE, i=l, 2, j=l, 2, k=l, 2; X _{1 3⁄4} X ₂ ,

X ₂ ', Xj", X ₂ ", ΧΓ, X ₂ "' are weight coefficients, and the specific determination method of the weight coefficient is prior art, and will not be described in detail herein.

 Assuming the performance characterization values of each HPE obtained by each sampling in the above table 401, the performance characterization values of each HPE corresponding to each sample determined by the analytic method are as follows:

Step 403: Determine, according to the performance characterization value of each cockroach corresponding to each sample, an analytic hierarchy process to determine a performance characterization value of the group to be controlled corresponding to each sample, specifically:

P2—into 1 ^x K ₂ i into ₂ ^X - ₂ 2 _?

Wherein, the performance characterization value of the HPE group to be controlled corresponding to the i-th sample is represented; ΧΓ' Χ ₂ "" Χ _ι Χ ₂ is a weight coefficient.

 It is assumed that the performance characterization value of each enthalpy corresponding to each sample determined by the above-mentioned step 402 is determined by the analytic hierarchy process. Show:

 Step 404: Determine the operating frequency value of each enthalpy and the group to be controlled by using multiple linear regression method according to the working frequency value of each cockroach obtained for each sample and the performance characterization value of the group to be controlled corresponding to each sampling. The correlation coefficient between the performance characterization values is:

Let F corpse (f„, f ₂₁ ), F ₂ =(f ₁₂ , f ₂₂ ), P=(pp ₂ ); where, denote the working frequency value obtained by sampling the working frequency of the jth HPE at the ith time ;

_{^ ΣFiFi Σ ^ Σ ^ 1χ1 +} 4Χ4 _ (1 + 4) χ (1 + 4) = 3

¹¹ ^ ^{1 1} η 2

_∑Fp , 3⁄4∑ ^ _{lx5 + 4x6} , (l ₊ 4)x(6 ₊ 5) ₌

l ^p ^ ¹ n 2

_{∑F2P ∑F 2 ∑P} , _{LX5 +} 3 _x6 ,(l ₊ 3)x(6 ₊ 5) ^

2 ^p ^ ² n 2

_{∑pP -∑} ∑^ _5x5十_6X6 -(5 ₊ 6)x _{(5 + 6)} = _05;

p ^p厶n 2

 Substituting the above calculated values into the following equations:

| ¹ ll ^b l ⁺¹ 12 ^b 2 ⁼¹ lp .

)l ₂₁ Vl ₂₂ b ₂ =l ₂ ' B, b

 29 29

Can characterize the correlation coefficient between values.

 The correlation coefficients between the final determined operating frequency values of each HPE and the performance characterization values of the HPE group to be controlled are shown in the following table:

 Step 405: Determine whether a difference between a current memory bandwidth of the heterogeneous processor and a theoretical memory bandwidth is greater than a preset bandwidth difference.

 Assume that the theoretical memory bandwidth of the heterogeneous processor is 16GB/s and the preset bandwidth difference is 2GB/s. If the current memory bandwidth of the heterogeneous processor is 15 GB/s, the difference between the current memory bandwidth of the heterogeneous processor and the theoretical memory bandwidth is not greater than the preset bandwidth difference, and the process proceeds to step 406;

 If the current memory bandwidth of the heterogeneous processor is 2 GB/s, the difference between the current memory bandwidth and the theoretical memory bandwidth of the heterogeneous processor is greater than the preset bandwidth difference, and the process proceeds to step 407.

 Step 406: Control power reduction of the two HPEs to release the memory bandwidth, and the power control process ends.

 Step 407: Determine, according to a correspondence between the preset coefficient interval and the working frequency value, a working frequency value corresponding to a coefficient region in which a correlation coefficient between each HPE operating frequency value and a performance representative value of the HPE group to be controlled is located.

 In Embodiment 2 of the present invention, the coefficient interval specifically includes three:

 Coefficient interval 1: an interval less than 0.5;

 Coefficient interval 2. An interval greater than or equal to 0.5 and less than 0.8;

 Coefficient interval 3, an interval greater than or equal to 0.8;

The working frequency value corresponding to the preset coefficient interval 1 is fl, and the operating frequency corresponding to the coefficient interval 2 The value is £2, and the operating frequency value corresponding to the coefficient interval 3 is β, where ί!<ί2< β. Because the correlation coefficient between the operating frequency value of HPE 1 and the performance characterization value of the group to be controlled is 0.82, which falls within the coefficient interval 3, the operating frequency value corresponding to HPE1 is f3; the operating frequency value of HPE2 and the HPE to be controlled The correlation coefficient between the performance characterization values of the group is 0.36, which falls within the coefficient interval 1, so the operating frequency value corresponding to HPE2 is fl.

 Step 408: Pre-determine whether the power of the heterogeneous processor exceeds the maximum power allowed by the heterogeneous processor after performing power control on each HPE according to the determined working frequency value.

 When it is determined that power control of each HPE is performed according to the determined operating frequency value, if the power of the heterogeneous processor does not exceed the maximum power allowed by the heterogeneous processor, proceed to step 409; when determining that each of the determined operating frequency values is determined After the HPE performs power control, if the power of the heterogeneous processor exceeds the maximum power allowed by the heterogeneous processor, the process proceeds to step 410.

 Step 409: Perform power control on each HPE according to the working frequency value determined in step 407, and the power control process ends.

 Step 410: Reduce the operating frequency value of each HPE determined in step 407, and perform power control on each HPE according to the reduced operating frequency value, and the power control process ends.

 Example 3:

 Based on the same inventive concept, the power control method according to the above embodiment of the present invention, and correspondingly, the embodiment of the present invention further provides a power control device. The structure of the device is as shown in FIG. 5, and the specific structure includes:

 The sampling unit 501 is configured to: perform performance indicators of each HPE in the HPE group to be controlled that includes at least two heterogeneous processing engines HPE in the heterogeneous processor, and perform the operating frequency of the HPE corresponding to the performance indicator at least once Sampling

 a first determining unit 502, configured to determine, according to a performance index value of each HPE obtained for each sample, a performance characterization value of the HPE group to be controlled corresponding to each sampling;

The second determining unit 503 is configured to determine, according to the working frequency value of each HPE obtained by each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling, the working frequency value of each HPE and the HPE group to be controlled. a correlation coefficient between performance characterization values; wherein, the correlation coefficient is larger, the table The greater the effect of the operating frequency value corresponding to the HPE on the performance characterization value of the HPE group to be controlled; the control unit 504 is configured to calculate a correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled, Each HPE is power controlled.

 Further, the first determining unit 502 is specifically configured to determine, according to the performance index value of each HPE obtained by each sampling, a performance characterization value of each HPE corresponding to each sampling; according to each HPE corresponding to each sampling The performance characterization value determines the performance characterization value of the HPE group to be controlled corresponding to each sample.

 Further, the first determining unit 502 is specifically configured to: for each HPE, when the performance index value of the sampled HPE is one, determine that the performance characterization value of the HPE corresponding to each sample is each sampling. The obtained performance index value of the HPE; when the sampled obtained HPE has the performance index value of at least two, the analytic hierarchy process, the network analysis method, or the multi-attribute performance decision method is used to obtain the species for each sample. At least two performance index values of the HPE are calculated to determine a performance characterization value of the HPE corresponding to each sample.

 Further, the first determining unit 502 is specifically configured to calculate, by using an analytic hierarchy process, a network analysis method, or a multi-attribute performance decision method, a performance characterization value of each HPE corresponding to each sample to determine a corresponding sample for each sample. The performance characterization value of the HPE group to be controlled.

 Further, the second determining unit 503 is specifically configured to use a multiple linear regression method, a multiple nonlinear regression method or a complex correlation coefficient method, and the working frequency value of each HPE obtained for each sampling and the corresponding control to be controlled for each sample The performance characterization values of the HPE group were calculated to determine the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 Preferably, the control unit 504 is further configured to determine, before power control of each HPE, that a difference between a current memory bandwidth and a theoretical memory bandwidth of the heterogeneous processor is greater than a preset bandwidth difference.

Further, the control unit 504 is specifically configured to determine, according to a preset relationship between the coefficient interval and the working frequency value, a coefficient of a correlation coefficient between an operating frequency value of each HPE and a performance representative value of the HPE group to be controlled. The operating frequency value corresponding to the interval; wherein, the larger the coefficient value in the coefficient interval, the larger the operating frequency value corresponding to the coefficient interval; and the power control for each HPE according to the determined working frequency value. Preferably, the control unit 504 is further configured to: before performing power control on each of the HPEs according to the determined operating frequency value, pre-determining the power of the heterogeneous processor after performing power control on each of the HPEs according to the determined operating frequency value. Does not exceed the maximum power allowed by the heterogeneous processor.

 The functions of the above units may correspond to corresponding processing steps in the flow shown in any of FIG. 2-4, and details are not described herein again.

 Example 4:

 Based on the same inventive concept, an embodiment of the present invention further provides a power control device, which is shown in FIG. 6 and includes a processor 601, a memory 602, and a communication bus 600. The communication bus 600 is used for each part of the device. Inter-connection communication; the processor 601 is configured to execute an executable module, such as a computer program, stored in the memory 602. The memory 602 may include a high speed random access memory (RAM: Random Access Memory) and may also include a non-volatile memory such as a disk memory.

 In some embodiments, the memory 602 stores a program 6021, which can be executed by the processor 601, and the program 6021 includes: for each of the to-be-controlled groups of the heterogeneous processor including at least two heterogeneous processing engines The performance indicator, and the working frequency of the 对应 corresponding to the performance indicator are at least once; determining the performance characterization value of the group to be controlled corresponding to each sampling according to the performance index value of each enthalpy obtained by each sampling; Determine the correlation coefficient between the operating frequency value of each enthalpy and the performance characterization value of the group to be controlled based on the operating frequency value of each enthalpy obtained for each sampling and the performance characterization value of the group to be controlled corresponding to each sampling. Wherein, the greater the correlation coefficient, the greater the effect of characterizing the corresponding operating frequency value of the corresponding ΗΡΕ group on the performance characterization value of the ΗΡΕ group; according to the operating frequency value of each ΗΡΕ and the performance characterization value of the ΗΡΕ group to be controlled Correlation coefficient, power control for each type.

 In some embodiments, the program 6021 specifically includes: determining, according to the performance index value of each ΗΡΕ obtained for each sample, a performance characterization value of each 对应 corresponding to each sampling; The performance characterization value determines the performance characterization value of the group to be controlled corresponding to each sample.

In some embodiments, the program 6021 specifically includes: for each ΗΡΕ, when sampled When the performance index value of the HPE is one, it is determined that the performance characterization value of the HPE corresponding to each sample is the performance index value of the HPE obtained by each sampling; and the performance index value of the HPE obtained by sampling For at least two types, at least two performance index values of the HPE obtained by each sampling are calculated by using an analytic hierarchy process, a network analysis method, or a multi-attribute performance decision method to determine the corresponding species for each sample. Performance characterization values for HPE.

 In some embodiments, the program 6021 specifically includes: using a analytic hierarchy process, a network analysis method, or a multi-attribute performance decision method to calculate a performance characterization value of each HPE corresponding to each sample to determine a corresponding sample for each sample. The performance characterization value of the HPE group to be controlled.

 In some embodiments, the program 6021 specifically includes: using a multiple linear regression method, a multiple nonlinear regression method, or a complex correlation coefficient method, and the operating frequency value of each HPE obtained for each sample and the corresponding control to be controlled for each sample The performance characterization values of the HPE group were calculated to determine the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled.

 In some embodiments, the program 6021 further includes: determining, before power control of each HPE, a difference between a current memory bandwidth and a theoretical memory bandwidth of the heterogeneous processor that is greater than a preset bandwidth difference.

 In some embodiments, the program 6021 specifically includes: determining, according to a preset relationship between the coefficient interval and the working frequency value, a correlation coefficient between an operating frequency value of each HPE and a performance representative value of the HPE group to be controlled. The operating frequency value corresponding to the coefficient interval; wherein, the larger the coefficient value in the coefficient interval, the larger the operating frequency value corresponding to the coefficient interval; and the power control for each HPE according to the determined working frequency value.

 In some embodiments, the program 6021 further includes: pre-determining the power of the heterogeneous processor after performing power control on each of the HPEs according to the determined operating frequency value, before performing power control on each of the HPEs according to the determined operating frequency value. Does not exceed the maximum power allowed by the heterogeneous processor.

Those skilled in the art will appreciate that embodiments of the present invention can be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or a combination of software and hardware. Moreover, the present invention may employ computer usable storage media (including but not limited to disk) in one or more of the computer usable program code embodied therein. The form of a computer program product implemented on a memory, CD-ROM, optical storage, etc.).

 The present invention has been described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (system), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowcharts and/or block diagrams, and combinations of flow and/or blocks in the flowcharts and/or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine for the execution of instructions for execution by a processor of a computer or other programmable data processing device. Means for implementing a particular function in a block or blocks of a flow or a flow and/or a block diagram of a flow diagram.

 The computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device. The apparatus implements a particular function in a block or blocks of a flow or a flow and/or block diagram of the flowchart.

 These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device. The instructions provide steps for implementing a particular function in a block or blocks of a flow or a flow and/or block diagram of a flowchart.

 Although the preferred embodiment of the invention has been described, it will be apparent to those skilled in the < Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and the modifications and modifications

 It is apparent that those skilled in the art can make various modifications and changes to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, it is intended that the present invention cover the modifications and modifications of the embodiments of the invention.

Claims

Rights request 1. A power control method, characterized by including: Sampling at least once the performance index of each HPE in the HPE group to be controlled that contains at least two heterogeneous processing engines HPE in the heterogeneous processor, and the operating frequency of the HPE corresponding to the performance index; Based on the performance index value of each HPE obtained from each sampling, the performance characterization value of the HPE group to be controlled corresponding to each sampling is determined; According to the working frequency value of each HPE obtained from each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling, determine the relationship between the working frequency value of each HPE and the performance characterization value of the HPE group to be controlled. Correlation coefficient; Wherein, the greater the correlation coefficient, the greater the effect of the operating frequency value representing the corresponding HPE on the performance representation value of the HPE group to be controlled; Power control is performed on each HPE based on the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled. 2. The method according to claim 1, characterized in that, based on the performance index value of each HPE obtained from each sampling, determining the performance characterization value of the HPE group to be controlled corresponding to each sampling includes: The performance index value of each HPE obtained by sampling is determined to determine the performance characterization value of each HPE corresponding to each sampling; According to the performance characterization value of each HPE corresponding to each sampling, the performance characterization value of the HPE group to be controlled corresponding to each sampling is determined. 3. The method of claim 2, characterized in that, based on the performance index value of each HPE obtained from each sampling, determining the performance characterization value of each HPE corresponding to each sampling includes: for each HPE, When the performance index value of the type of HPE obtained by sampling is one, the performance characterization value of the type of HPE corresponding to each sampling is determined to be the performance index value of the type of HPE obtained by each sampling; When there are at least two performance index values of the type of HPE obtained by sampling, use the analytic hierarchy process, network analysis method or multi-attribute performance decision-making method to measure at least two types of HPE obtained by each sampling. The performance index value is calculated to determine the performance characterization value of the HPE corresponding to each sampling. 4. The method according to claim 2, characterized in that, based on the performance characterization value of each HPE corresponding to each sampling, determining the performance characterization value of the HPE group to be controlled corresponding to each sampling includes: using hierarchical analysis Method, network analysis method or multi-attribute performance decision-making method is used to calculate the performance characterization value of each HPE corresponding to each sampling to determine the performance characterization value of the HPE group to be controlled corresponding to each sampling. 5. The method according to any one of claims 1 to 4, characterized in that, based on the working frequency value of each HPE obtained from each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling, each HPE is determined. The correlation coefficient between the operating frequency value of HPE and the performance characterization value of the HPE group to be controlled includes: using multiple linear regression method, multiple nonlinear regression method or complex correlation coefficient method, for each HPE obtained from each sampling The working frequency value and the performance characterization value of the HPE group to be controlled corresponding to each sampling are calculated to determine the correlation coefficient between the working frequency value of each HPE and the performance characterization value of the HPE group to be controlled. 6. The method according to any one of claims 1 to 5, characterized in that, before performing power control on each HPE, it further includes: Determine that the difference between the current memory bandwidth and the theoretical memory bandwidth of the heterogeneous processor is greater than the preset bandwidth difference. 7. The method according to any one of claims 1 to 6, characterized in that, power control is performed on each HPE according to the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled, include: According to the corresponding relationship between the preset coefficient interval and the operating frequency value, determine the operating frequency value corresponding to the coefficient interval in which the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled is located; where, the coefficient The greater the coefficient value in the interval, the greater the operating frequency value corresponding to the coefficient interval; According to the determined operating frequency value, power control is performed for each HPE. 8. The method according to claim 7, characterized in that, before performing power control on each HPE according to the determined operating frequency value, it further includes: After pre-judgment and power control for each type of HPE based on the determined operating frequency value, the power of the heterogeneous processor does not exceed the maximum allowable power of the heterogeneous processor. 9. A power control device, characterized in that it includes: The sampling unit is used to measure at least once the performance index of each HPE in the HPE group to be controlled including at least two heterogeneous processing engines HPE in the heterogeneous processor, and the working frequency of the HPE corresponding to the performance index. Sample; The first determination unit is used to determine the performance characterization value of the HPE group to be controlled corresponding to each sampling based on the performance index value of each HPE obtained from each sampling; The second determination unit is used to determine the working frequency value of each HPE and the HPE group to be controlled based on the working frequency value of each HPE obtained from each sampling and the performance characterization value of the HPE group to be controlled corresponding to each sampling. Correlation coefficient between performance characterization values; wherein, the greater the correlation coefficient, the greater the effect of the operating frequency value representing the corresponding HPE on the performance characterization value of the HPE group to be controlled; The control unit is used to control the power of each HPE according to the correlation coefficient between the operating frequency value of each HPE and the performance characterization value of the HPE group to be controlled. 10. The device according to claim 9, wherein the first determination unit is specifically configured to determine the performance of each HPE corresponding to each sampling according to the performance index value of each HPE obtained from each sampling. Characterization value; According to the performance characterization value of each HPE corresponding to each sampling, the performance characterization value of the HPE group to be controlled corresponding to each sampling is determined. 11. The device according to claim 10, wherein the first determination unit is specifically configured to determine each HPE when the sampled performance index value of the HPE is one. The performance characterization value of the type of HPE corresponding to the sampling is the performance index value of the type of HPE obtained by each sampling; when there are at least two performance index values of the type of HPE obtained by sampling, the analytic hierarchy process, network analysis method or The multi-attribute performance decision-making method calculates at least two performance index values of the type of HPE obtained from each sampling to determine the performance characterization value of the type of HPE corresponding to each sampling. 12. The device according to claim 10, characterized in that the first determination unit is specifically configured to use the analytic hierarchy process, the network analysis method or the multi-attribute performance decision-making method to determine the performance of each HPE corresponding to each sampling. Performance characterization values are calculated to determine the performance of the HPE group to be controlled corresponding to each sampling. Performance characterization values. 13. The device according to any one of claims 9 to 12, characterized in that the second determination unit is specifically configured to adopt a multiple linear regression method, a multiple nonlinear regression method or a complex correlation coefficient method, for each The working frequency value of each HPE sampled and the performance characterization value of the HPE group to be controlled corresponding to each sampling are calculated to determine the correlation coefficient between the working frequency value of each HPE and the performance characterization value of the HPE group to be controlled. . 14. The device according to any one of claims 9-13, characterized in that the control unit is also used to determine the current memory bandwidth and the theoretical memory bandwidth of the heterogeneous processor before performing power control on each HPE. The difference between them is greater than the preset bandwidth difference. 15. The device according to any one of claims 9 to 14, characterized in that the control unit is specifically configured to determine the working frequency value of each HPE according to the corresponding relationship between the preset coefficient interval and the working frequency value. The working frequency value corresponding to the coefficient interval in which the correlation coefficient between the performance characterization values of the HPE group to be controlled is; wherein, the greater the coefficient value in the coefficient interval, the greater the operating frequency value corresponding to the coefficient interval; According to the determined operating frequency value, perform power control for each HPE. 16. The device according to claim 15, wherein the control unit is further configured to pre-judge the power control of each type of HPE based on the determined operating frequency value before performing power control on each type of HPE. After HPE performs power control, the power of the heterogeneous processor does not exceed the maximum power allowed by the heterogeneous processor.