A parameter optimization method and a parameter optimization device

Abstract

The application provides a parameter optimization method and a parameter optimization device. The parameter optimization method comprises constructing a test set and constructing a Gaussian process model based on an anisotropic Gaussian kernel function; obtaining a plurality of parameter groups of interest of a target computer, inputting the parameter groups of interest into the Gaussian process model, determining a predicted mean and a predicted variance; determining an expected improvement value based on the predicted mean and the predicted variance, and taking the parameter group of interest corresponding to the expected improvement value meeting a preset condition as a candidate parameter group; performing a hardware performance test on the candidate parameter group; updating the Gaussian process model based on an updated test set, and taking the current candidate parameter group as an optimized parameter group in the case that a test result determined based on the current candidate parameter group meets a convergence condition. In this way, the time consumption of parameter optimization can be greatly reduced, and the parameter optimization efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of computer hardware performance optimization technology, specifically to a parameter optimization method and a parameter optimization device. Background Technology

[0002] Turbo Boost technology is key to improving the dynamic performance of laptops. PL2 and PL4, as core power management constraints of the CPU, directly determine the computer's short-term high load and instantaneous peak performance. The proper configuration of these two parameters is crucial for balancing the overall performance, hardware stability, and energy efficiency of the laptop.

[0003] The current mainstream method in the industry is to optimize PL2 and PL4 parameters using full-parameter space enumeration. This method requires hardware testing of a large number of parameter combinations, resulting in a long overall optimization time (over 10 hours) and extremely low testing efficiency. Repeated high-load testing can also accelerate hardware aging. Furthermore, this method cannot reuse historical optimization experience for new models, leading to significant uncertainty in initial exploration. In addition, dangerous parameter combinations are prone to occur during testing, posing risks of hardware damage such as power supply overload and thermal shock.

[0004] Currently, there is an urgent need for an intelligent parameter optimization method to address the problems of extremely low optimization efficiency, hardware security risks, and inability to balance performance and energy efficiency associated with traditional methods. Summary of the Invention

[0005] This application addresses the aforementioned technical problems in the existing technology. The purpose of this application is to provide a parameter optimization method and apparatus that can achieve intelligent iterative optimization of parameters, significantly reducing the number of tests and optimization time, while balancing computer performance and energy efficiency, avoiding hardware damage risks, and improving the efficiency, accuracy, and safety of parameter optimization.

[0006] According to the first aspect of this application, a parameter optimization method is provided, the parameter optimization method comprising: Hardware performance tests are conducted based on preset parameter sets of the target computer to determine the test results corresponding to each preset parameter set in order to construct a test set. Using the preset parameter set as input and the corresponding test results as observation labels, a Gaussian process model is constructed by combining the anisotropic Gaussian kernel function; Multiple sets of parameters of interest are obtained for the target computer, and the sets of parameters of interest are input into the Gaussian process model to determine the prediction mean and prediction variance of each set of parameters of interest. Based on the predicted mean and predicted variance, the expected improvement value corresponding to each group of attention parameters is determined, and the group of attention parameters corresponding to the expected improvement value that meets the preset conditions is taken as the candidate parameter group. Hardware performance tests are performed on the candidate parameter group to determine the test results corresponding to the candidate parameter group and update them to the test set; The Gaussian process model is updated based on the updated test set. If the test results determined based on the current candidate parameter set meet the convergence condition, the current candidate parameter set is used as the optimization parameter set.

[0007] According to a second aspect of this application, a parameter optimization apparatus is provided, the parameter optimization apparatus including a processor configured to execute the steps of the parameter optimization method described in various embodiments of this application.

[0008] According to a third aspect of this application, a computer program product is provided, the computer program product comprising computer-executable instructions, which, when executed by a processor, implement the steps of the parameter optimization method described in various embodiments of this application.

[0009] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: This application provides a parameter optimization method that first constructs an initial test set through testing with a preset parameter set, and then constructs a Gaussian process model using an anisotropic Gaussian kernel function. The Gaussian process model is then used to quickly predict the mean and variance of a large number of parameters of interest, and based on this, the desired improvement values ​​are determined, and candidate parameter sets that meet preset conditions are selected. Hardware performance testing is then performed on the selected small number of candidate parameter sets. This significantly reduces the number of hardware test sets, shortens the overall parameter optimization time, and avoids the equipment aging problem caused by full enumeration testing. Simultaneously, the test set is updated based on the test results obtained from hardware performance testing using the candidate parameter sets, and the Gaussian process model is updated based on the updated test set to continuously optimize the candidate parameter selection accuracy until the convergence condition is met, at which point the optimized parameter set is output.

[0010] By screening high-potential candidate parameter sets for hardware performance testing, the large number of repetitive tests caused by traditional full parameter enumeration is avoided, which helps improve the parameter optimization efficiency of the target computer. Simultaneously, the Gaussian process model is dynamically and iteratively updated based on the actual test results obtained from hardware performance testing of candidate parameter sets. This continuously corrects the prediction bias of the Gaussian process model regarding parameter performance, adjusts the parameter optimization direction in real time, and prevents the optimization process from getting trapped in local optima prematurely. This allows for rapid convergence and determination of the optimal parameter set for the target computer, effectively improving the matching and adaptability between the optimized parameter set and the overall hardware operating conditions.

[0011] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above description and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0012] In drawings that are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. Similar reference numerals with different letter suffixes may indicate different examples of similar components. The drawings generally illustrate various embodiments by way of example rather than limitation, and are used together with the specification and claims to illustrate the disclosed embodiments. Such embodiments are illustrative and exemplary, and are not intended to be exhaustive or exclusive embodiments of the method, apparatus, system, or non-transitory computer-readable medium having instructions for implementing the method.

[0013] Figure 1 A flowchart of a parameter optimization method according to an embodiment of this application is shown.

[0014] Figure 2 Another flowchart of a parameter optimization method according to an embodiment of this application is shown.

[0015] Figure 3 A parameter optimization apparatus according to an embodiment of this application is shown. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific examples, but these are not intended to limit the scope of this application.

[0017] The terms "first," "second," and similar words used in this application do not indicate any order, quantity, or importance, but are merely used for distinction. The terms "including" or "comprising," etc., used in this application mean that the element preceding the word encompasses the elements listed after the word, and do not exclude the possibility of encompassing other elements. In this application, the arrows shown in the figures for each step are merely examples of the execution order, not limitations. The technical solution of this application is not limited to the execution order described in the embodiments. The steps in the execution order can be combined, broken down, or rearranged, as long as the logical relationship of the executed content is not affected.

[0018] All terms used in this application (including technical or scientific terms) have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein. Technologies and equipment known to one of ordinary skill in the art may not be discussed in detail, but where appropriate, such technologies and equipment should be considered part of the specification.

[0019] Figure 1 A flowchart of a parameter optimization method according to an embodiment of this application is shown. In this application, the arrows shown in the figure for each step are merely examples of the execution order and not limitations. The technical solution of this application is not limited to the execution order described in the embodiments. The steps in the execution order can be combined, decomposed, or rearranged, as long as the logical relationship of the execution content is not affected.

[0020] In step S101, hardware performance testing is performed based on the preset parameter group of the target computer to determine the test results corresponding to each preset parameter group in order to construct a test set.

[0021] The preset parameter group and the candidate parameter group mentioned below include the short-term turbo power limit PL2 and the peak turbo power limit PL4. PL2 is the highest power limit that the CPU can reach in a short period of time, which directly determines the turbo performance of the target computer in a short-term high-load scenario; PL4 is the maximum power limit of the CPU during instantaneous bursts, which directly determines the peak turbo performance of the target computer under instantaneous high computing demands.

[0022] The parameter optimization methods described in the various embodiments of this application are all intelligent optimization methods for the CPU turbo frequency power consumption parameters of the target computer, and the selection of the optimal parameter combination of PL2 and PL4 that is adapted to the overall hardware performance.

[0023] In some embodiments of this application, the value range of each parameter in the parameter group is determined based on the hardware operating constraints of the target computer.

[0024] This range of values ​​defines the effective range for subsequent parameter selection, testing, and iterative optimization. Parameter combinations outside this range are invalid. By limiting subsequent parameter selection and hardware performance testing to a safe range, invalid testing and hardware degradation are avoided, while further improving parameter optimization efficiency.

[0025] Specifically, the hardware operating constraints include current constraints, transient temperature rise constraints, and power limit constraints. The current constraint limits the maximum value of PL2 based on voltage and continuous current, for example, based on the thermal design current of the power supply module. T dc and processor core voltage V core By determining the maximum value of PL2 and limiting the maximum power consumption of short-term turbo boost, overload and overheating caused by long-term high-current operation of the power supply circuit can be prevented.

[0026] The transient temperature rise constraint, based on the difference between PL4 and PL2, limits the instantaneous temperature rise of the hardware to a threshold temperature. This helps to avoid sudden thermal shocks caused by peak power consumption and protects the heat dissipation module and CPU chip. The threshold temperature can be 10°C, and is not limited thereto.

[0027] The power upper limit constraint limits PL2 to be less than PL4, which helps to ensure that the short-term power consumption limit is lower than the instantaneous peak power consumption. Furthermore, the value of PL4 is not greater than the smaller of the battery discharge power limit and the rated peak power, which helps to constrain the instantaneous peak power consumption within a safe range, prevent battery overload discharge and power supply circuit overload operation, and reduce the risk of hardware damage.

[0028] For example, the value ranges of PL2 and PL4 in the parameter group can be determined according to the following formula (1):

[0029] In formula (1), T dc This indicates the thermal design current (e.g., 140A). V core This indicates the core voltage (e.g., 1.2V). ΔT This represents the transient temperature rise. 0.85 is a preset safety value, while 0.05 and 4.5 are hardware thermal characteristic fitting coefficients determined based on historical experience. In other implementations, the safety value and hardware characteristic fitting coefficients can be reconfigured; no specific limitations are imposed on this. Battery discharge This represents the maximum discharge power that the target computer can stably output under full load.

[0030] Thus, the range of values ​​for PL2 and PL4 can be determined using formula (1).

[0031] It should be noted that formula (1) is only used as an example and does not constitute a limitation on specific solutions.

[0032] For example, the values ​​in formula (1) can be modified based on historical experience. For instance, the threshold temperature of 10℃ can be replaced with 11℃, and the coefficient of 0.9 can be replaced with 0.8. Alternatively, the constraint formulas for PL2 and PL4 can be set based on other hardware operation constraints.

[0033] In some embodiments of this application, the preset parameter group includes a first parameter group and a second parameter group. A hardware feature vector is determined based on the hardware operating parameters of the target computer. The first parameter group is determined based on the hardware feature vector through multiple linear regression. The hardware operating parameters include heat dissipation parameters, power supply parameters, and maximum power consumption parameters.

[0034] Among them, the hardware feature vector can be understood as a one-dimensional feature array obtained by quantizing the hardware operating parameters of the target computer on a unified scale. It is a one-dimensional feature array formed by orderly splicing together the hardware operating parameters of different dimensions and numerical magnitudes through standardized scale alignment transformation, eliminating the differences in the dimensions and numerical magnitudes of each indicator, converting them into quantized feature values ​​under the same numerical range.

[0035] In some embodiments, the hardware feature vector is obtained by transforming the hardware operating parameters. This may include numerical scaling transformations such as linear scaling, compression, or stretching of numerical ranges for the original hardware operating parameters with different ranges and dimensions; nonlinear mapping transformations such as logarithmic, exponential, or piecewise functions are used for indicators where the hardware operating parameters are nonlinearly correlated; and corresponding weights can be configured for different hardware operating parameters, and then normalization operations are performed in combination with preset specification parameters.

[0036] This is provided as an example only and does not constitute a limitation on any specific solution.

[0037] In this embodiment, the heat dissipation parameters may include the number of heat pipes and fan airflow (CFM) of the target computer, and the power supply parameters may include the number of phases of the power supply module (VRM) and the thermal design current (TDC). T dc The maximum power consumption parameter includes the CPU's maximum continuous power consumption. TDP max .

[0038] In some embodiments, coupled calculations can be performed based on the number of heat pipes and CFM, and scaling can be applied to obtain a first-dimensional feature value that characterizes the overall heat dissipation capability of the system.

[0039] For example, the first dimension feature value = The number of heat pipes multiplied by the CFM (Coefficient of Motion) is a rough indicator of the absolute cooling capacity of a system; the more heat pipes and the larger the CFM, the better the cooling performance. Scaling allows the values ​​to fall within a small range (e.g., 0.5-2.0), making them comparable to the numerical scales of other dimensions. This results in more stable and easier-to-converge multiple linear regression calculations.

[0040] In some embodiments, it can be based on the number of VRM phases and T dc After weighted ratio conversion and normalization, the second-dimensional characteristic value representing the redundancy carrying capacity of the whole machine's power supply is obtained.

[0041] For example, the second dimension feature value = In this context, VRM phase number represents the number of power supply phases in the power supply module. The more power supply phases there are, the stronger the power supply module's ability to instantaneously carry large currents. T dc For the thermal design current of the power supply module, divide by T dc The calculation is standardized, and the result represents the redundancy of the power supply phases per unit current. The higher the value, the more sufficient the power supply module margin is, making it more suitable for high-power turbo boost conditions. During the calculation, the number of VRM phases is multiplied by a coefficient of 10 beforehand, which is a numerical scaling operation used to adjust the range of index values, so that the second dimension feature value is on a similar numerical scale to other dimension feature values, ensuring the stability and easy convergence of subsequent multiple linear regression calculations.

[0042] In some embodiments, the maximum continuous power consumption of the CPU can be used as a basis. TDP max A base-10 logarithmic transformation is performed to obtain the third-dimensional feature value used to characterize the processor's power consumption release potential.

[0043] For example, the third-dimensional feature value = log 10 ( TDP max ),in, TDP max This represents the maximum sustained power consumption allowed by both the CPU itself and the target computer system design, characterizing the upper limit of the processor's basic power consumption. Since there is no linear correlation between processor power consumption and overall system performance, but rather a logarithmic correlation, it is important to understand... TDP max Using a base-10 logarithm for non-linear mapping can better reflect the performance variation patterns of the hardware itself, accurately capture the differences in power consumption potential between different models, and simultaneously achieve scale adaptation of this third-dimensional feature value, unifying the value range with other dimensional feature values.

[0044] In some embodiments, the hardware feature vector V=[ log 10 ( TDP max )).

[0045] For example, a target computer has 5 heat pipes, 27 CFM, and 8 VRM phases. T dc =140A、 TDP max =120W, then the first feature value = 5×27 / 100 = 1.35, the second feature value = 8×10 / 140≈0.57, the third feature value ≈2.08, and the hardware feature vector V = [1.35, 0.57, 2.08].

[0046] This is merely an illustrative example and does not constitute a limitation on any specific solution.

[0047] In the embodiments of this application, the hardware feature vector V can be input into a multiple linear regression model to obtain the first parameter set, wherein the multiple linear regression model is as shown in formula (2): [ Formula (2) = W × V + b Among them, in formula (2) PL 2init This represents the initial value of PL2. PL 4init Let W represent the initial value of PL4, W be the weight matrix, and b be the bias vector. W and b can be determined by training based on historical data of historical computer models.

[0048] For example, a large number of hardware feature vectors of different computer models can be obtained in advance, as well as the optimal PL2 and PL4 parameters of the corresponding models after debugging, as a sample dataset. The multiple linear regression model can be trained offline to solve for the weight matrix W and the bias vector b.

[0049] Then, the hardware feature vector V is input into the trained multiple linear regression model, and the initial values ​​of PL2 and PL4 that are adapted to the carrying capacity of the target computer hardware are calculated through matrix linear operations. Then, combined with the range constraints of PL2 and PL4, the first parameter set is obtained.

[0050] Thus, based on the parameter optimization method provided in this application embodiment, the initial values ​​of PL2 and PL4 adapted to the target computer model can be obtained quickly. The initial values ​​of PL2 and PL4 of the first parameter group are close to the potential optimal solution area, eliminating the need for blind testing from scratch, greatly reducing the number of blind hardware tests in the early stage, and helping to improve the screening efficiency.

[0051] For example, when W is b is In the case of W and b, substitute W and b into formula (2), and based on the above hardware feature vector V=[1.35,0.57,2.08], the initial values ​​of PL2 and PL4 are 70.42W and 153.92W, respectively.

[0052] By taking integer values, the initial value of PL2 is determined to be 70W and the initial value of PL4 is determined to be 150W. Thus, the first parameter set (70W, 150W) can be obtained.

[0053] In some embodiments of this application, representative values ​​are selected from the range of values ​​for each parameter as the second parameter group.

[0054] Specifically, the second parameter group can be obtained by selecting the boundary representative value within the range of values ​​of each parameter.

[0055] For example, based on the respective value ranges of PL2 and PL4, representative boundary values ​​can be selected from the interval boundary positions as representative values.

[0056] For example, assuming the range of PL2 is [40W, 90W] and the range of PL4 is [90W, 160W], we can take the minimum value of PL2 (40W) and the minimum value of PL4 (90W) to obtain a second set of parameters (40W, 90W); we can take the minimum value of PL2 (40W) and the maximum value of PL4 (160W) to obtain another set of second parameters (40W, 160W); we can also take the maximum value of PL2 (90W) and the minimum value of PL4 (90W) to obtain yet another set of second parameters (90W, 90W).

[0057] This is merely an illustrative example and does not constitute a limitation on any specific solution.

[0058] In some embodiments of this application, the preset parameter group includes a first parameter group and at least one second parameter group obtained based on boundary representative values.

[0059] Boundary values ​​represent the critical limits of power consumption that the hardware can withstand. They allow for early verification of the hardware's performance under extreme power conditions, revealing shortcomings in heat dissipation and power supply under extreme loads. This prevents the model from recommending dangerous parameter combinations outside the safe range during subsequent iterations. Further processing based on this preset parameter set enriches the sample distribution of the initial test set, enabling the anisotropic Gaussian kernel Gaussian process model to fully learn the mapping relationships between parameters and overall system performance, temperature rise, etc., across the entire power consumption range from minimum to maximum. This effectively reduces the errors in the mean and variance of the Gaussian process model's early predictions, improves the accuracy of candidate parameter selection in subsequent iterations, accelerates the convergence speed of the optimal PL2 and PL4 parameter sets, and further enhances the overall efficiency of the parameter optimization process.

[0060] In this embodiment of the application, the hardware performance test can be understood as configuring each set of preset parameters to the target computer, running it under a unified environment and load conditions, and synchronously collecting the test data corresponding to each set of preset parameters to obtain real and objective hardware operation data.

[0061] In some other embodiments of this application, the hardware performance test includes: after loading the corresponding parameter group on the target computer, running a performance test program and synchronously collecting test data.

[0062] Specifically, the set of PL2 and PL4 parameters to be tested is written into the underlying power control module of the target computer. After loading, the device will run according to the short-term turbo frequency power consumption and peak power consumption limits of the set of parameters, ensuring that the power consumption constraints of the whole machine are completely matched with the set of parameters to be tested when the subsequent performance test program is running.

[0063] The performance testing program is a standardized software tool that automatically and continuously applies a stable load. Running on the target computer system, it applies a fixed computational pressure to the CPU. During operation, the program continuously consumes CPU resources, bringing the device to full or high load status, thereby triggering PL2 and PL4 power limiting mechanisms to ensure that heat dissipation, power supply, and power consumption meet real-world operating conditions. Simultaneously, it records various test data such as temperature, power consumption, frequency, and frequency throttling times, ensuring complete uniformity in test environment, load intensity, and runtime across different parameter groups, thus providing objective comparative results for each parameter group.

[0064] The performance testing program includes standardized benchmarking software, custom loop calculation scripts, and a comprehensive load testing suite for the entire machine. It can stably and continuously output controllable computing power load, and also has a built-in data acquisition function to synchronously record raw test data such as temperature, power consumption, frequency, and frequency reduction times, ensuring that the collected test data can truly reflect the actual operating status of the entire machine under the set of parameters.

[0065] In other embodiments of this application, test results are determined based on the test data to characterize the overall performance of the target computer.

[0066] The test results can be a comprehensive performance score, such as an analysis of multiple measured indicators, including CPU computing power, full-load temperature, power supply fluctuation range, and actual power consumption, to obtain a comprehensive performance score for evaluating the compatibility of the preset parameter group with the target computer hardware.

[0067] For example, the collected test data of different dimensions such as CPU computing power, full load temperature, actual power consumption, frequency of frequency reduction, and fan speed can be normalized. The weights of each indicator are assigned according to their impact on the overall stability, computing efficiency, and user experience, and the comprehensive performance score is obtained. This score serves as the quantitative test result of the overall operating performance of the machine under the corresponding parameter group.

[0068] As a preferred approach, the test set includes each preset parameter group and the corresponding comprehensive performance score.

[0069] In some embodiments of this application, the test data includes at least the CPU sustained high load performance score, daily application scenario performance score, and average power consumption throughout the test.

[0070] For example, when running performance testing programs for a target computer model, professional testing tools such as Cinebench R23 and PCMark 10 can be used. During the test, the average power consumption (AvgPower) of the entire machine is monitored simultaneously, and various test data are recorded in full. Finally, three types of test indicators are obtained: CPU sustained high load performance score (CB23), which reflects the CPU's sustained high load performance; daily application scenario performance score, which represents the smoothness of daily software use (PCMark); and average power consumption (AvgPower) during the full load operation phase of the entire machine.

[0071] Among them, the CPU sustained high load performance score reflects the CPU's turbo boost performance under long-term high load scenarios (such as video editing, code compilation, and running large software); the daily application scenario performance score reflects the CPU's instantaneous response and turbo boost performance under daily light and medium load scenarios such as web browsing, office software operation, and video conferencing; the average power consumption throughout the test reflects the overall power consumption level of the CPU under the test load when running with the parameter combination.

[0072] In this embodiment, the parameter group loaded on the target computer can be the preset parameter group mentioned above, the candidate parameter group to be described below, or other parameter groups that need to be tested for hardware performance. This embodiment is only used to illustrate the overall logic of testing hardware performance based on parameter groups and evaluating comprehensive performance scores based on hardware performance test results, and is not limited to a single operation step.

[0073] In some embodiments of this application, the CPU sustained high load performance score and the daily application scenario performance score are weighted and summed to obtain a weighted performance score; The average power consumption is weighted to obtain a weighted power consumption penalty term; The test results are determined based on the weighted performance score and the weighted power consumption penalty term.

[0074] Specifically, the CPU sustained high load performance score and daily application scenario performance score are weighted and summed based on preset weights to obtain a weighted performance score; the average power consumption is weighted based on preset weights to obtain a weighted power consumption penalty term; and the comprehensive performance score is obtained based on the weighted performance score and the weighted power consumption penalty term.

[0075] The preset weights are not specifically limited and can be set according to the usage scenario of the target computer.

[0076] In the process of evaluating comprehensive performance scores, average power consumption is used as a weighted power consumption penalty item. The higher the average power consumption, the larger the value of the weighted power consumption penalty item. The preset weight corresponding to the average power consumption can be set to a negative number to avoid the algorithm from selecting high-performance but high-power and high-heat parameter groups, thereby realizing the quantification of energy efficiency constraints.

[0077] For example, if sustained performance is prioritized for the target computer, a higher preset weight can be configured for the CPU sustained high load performance score (CB23). For instance, a preset weight of 0.7 can be set for CB23, 0.3 for PCMark, and -0.1 for AvgPower. This means that daily performance is important but has a slightly lower priority than sustained performance requirements. In this case, the overall performance score f(x) can be set to 0.7 × CB23 + 0.3 × PCMark. 0.1×AvgPower.

[0078] This is merely an illustrative example and does not constitute a limitation on any specific solution.

[0079] Returning to the embodiment of this application, in step S102, the preset parameter group is used as input, the corresponding test result is used as the observation label, and a Gaussian process model is constructed by combining the anisotropic Gaussian kernel function.

[0080] Specifically, each set of preset parameters in the test set is used as the input feature vector of the model, and the test results corresponding to each set of preset parameters are used as the observation labels of the model. Based on all input feature vectors and corresponding observation label samples, an anisotropic Gaussian kernel is used as the covariance function to construct a Gaussian process model.

[0081] The Gaussian process model measures the correlation similarity between any two sets of input samples using a kernel function. The anisotropic Gaussian kernel assigns different length scales to each dimension of the input feature vector, differentially characterizing the sensitivity of different input feature vectors to the output observation label, thereby more accurately fitting the nonlinear mapping relationship between the input feature vector and the test result. After training, the model can output the predicted mean and predicted variance for any input sample to be predicted.

[0082] As one specific embodiment, a test set of training samples is constructed, which includes the set of input parameters X. train With output performance set Y train Input parameter set X train Each element in the vector is an input feature vector representing a set of PL2 and PL4 parameter groups, for example, X train =[(80,170),(85,175),(90,180),……]. Here, the total number of all preset parameter sets used to complete the hardware performance test can be denoted as N. N is at least 2; for example, for the initial test set, N can be set to 2. Input parameter set X train The CCP contains N input feature vectors. The output performance set Y train Each element in the table represents the overall performance score corresponding to the preset parameter set. The overall performance score can be calculated using f(x) = 0.7 × CB23 + 0.3 × PCMark. Calculated using 0.1×AvgPower, for example, Y train =[0.85,0.88,0.82, ……].

[0083] In other embodiments of this application, different length scales are set for different parameters in the preset parameter group to obtain anisotropic Gaussian kernel functions; the similarity between each preset parameter group is determined using the anisotropic Gaussian kernel functions, and a covariance matrix is ​​constructed based on the similarity; a Gaussian process model that can output the predicted mean and predicted variance is trained based on the covariance matrix.

[0084] Specifically, the anisotropic Gaussian kernel function configures different length scales for the two independent parameter dimensions, PL2 and PL4, which can distinguish the difference in the sensitivity of the two parameters to the overall performance of the machine. When a small change in PL2 will cause a significant change in the overall performance score, its corresponding length scale value is smaller; if a large adjustment in the value of PL4 will cause a significant fluctuation in performance, its length scale value is larger.

[0085] In a preferred embodiment, the Gaussian process model employs an anisotropic Gaussian kernel function, which sets differentiated length scales based on the sensitivity differences of PL2 and PL4 to the overall performance, which is beneficial to significantly accelerate convergence.

[0086] Among them, PL2 affects the CPU's sustained high-load performance. Even a small change (±5W) can cause significant fluctuations in the overall system performance. Therefore, a smaller PL2 length scale can be set. PL4 only affects instantaneous peak performance and has a more gradual impact on overall performance. Therefore, a larger PL4 length scale can be set.

[0087] Specifically, the anisotropic Gaussian kernel function is shown in formula (3):

[0088] in, lPL 2 indicates the length of PL2. lPL 4 represents the length dimension of PL4, and, lPL 4 greater than lPL 2. k ( x i , x j The ) indicates the similarity of the overall performance scores between the two sets of preset parameters. x i Indicates the preset parameter group PL 2 i and PL 4 i ,Right now x i =( PL 2 i , PL 4 i ); x j Indicates the preset parameter group PL 2 j and PL 4 j ,Right now x j =( PL 2 j , PL 4 j ). For example, with lPL 2=5, lPL 4 = 10 x i =(80,170) x j Let's take (85, 175) as an example for calculation. k ( x i , x j =0.535.

[0089] As one embodiment, based on the anisotropic Gaussian kernel function, any two sets of preset parameter groups are calculated using the anisotropic Gaussian kernel function. x i =( PL 2 i , PL 4 i )and x j =( PL 2 j , PL 4 j Similarity between k ( x i , x j A covariance matrix of dimension N×N is constructed based on the similarity between each preset parameter group in the test set. K The elements in the matrix satisfy K [ i , j ]= k (X train [ i ] , X train [ j The covariance matrix possesses symmetric properties and satisfies... K [ i , j ]= K [ j , i Furthermore, the diagonal elements of the covariance matrix are always 1, which can fully represent the similarity relationship between all training parameter samples.

[0090] For the set of parameters to be predicted The similarity between the set of parameters of interest and each set of training samples in the test set is calculated using an anisotropic Gaussian kernel function, and a column vector is generated. Simultaneously, the similarity of the target parameter group itself is calculated. Based on the covariance matrix K Similarity vectork ( x i , x j and output performance set Y train It can be based on the formula Calculate the predicted mean μ, where, K -1 Represents the covariance matrix K The inverse matrix.

[0091] According to the formula The prediction variance σ is calculated to train the Gaussian process model. The trained Gaussian process model outputs a normally distributed N(μ, σ) pattern. 2 The prediction results can be used to simultaneously output the predicted mean μ and predicted variance σ for any group of parameters to be tested.

[0092] In other words, in step S103, multiple sets of parameters of interest for the target computer are obtained, the sets of parameters of interest are input into the Gaussian process model, and the predicted mean and predicted variance of each set of parameters of interest are determined.

[0093] Specifically, based on the hardware constraints of the target computer, the value ranges of parameters PL2 and PL4 in the parameter group can be determined. Based on these value ranges, a traversal step size can be set. Multiple sets of parameters of interest are then discretized based on the value range boundaries and the traversal step size. The traversal step size can be adjusted according to the optimization accuracy requirements; a smaller step size results in a larger number of parameter sets and a finer granularity of parameter search.

[0094] For example, assuming that the value range of PL2 is 60W to 100W and the value range of PL4 is 150W to 200W, and the traversal step size is set to 1W, 41 discrete values ​​can be generated based on the PL2 dimension and 51 discrete values ​​can be generated based on the PL4 dimension. The total number of parameter groups is 41×51=2091.

[0095] Each set of parameters of interest is sequentially input as the input sample to be predicted into the trained Gaussian process model. The Gaussian process model uses an anisotropic Gaussian kernel to solve for the similarity vector between the current set of parameters of interest and all N measured training samples. Currently focusing on the similarity of the parameter group itself. Then, combined with the inverse of the covariance matrix K -1 With output performance set Y train The predicted mean value for each group of parameters of interest is calculated sequentially. m With prediction variance s .

[0096] This is provided as an example only and does not constitute a limitation on any specific solution.

[0097] In step S104, the expected improvement value corresponding to each group of attention parameters is determined based on the predicted mean and predicted variance, and the group of attention parameters corresponding to the expected improvement value that meets the preset conditions is taken as the candidate parameter group.

[0098] Specifically, the predicted mean and predicted variance can be used as inputs to quantify the optimization potential of each set of focus parameters through the expected improvement function, which is a Bayesian optimized acquisition function.

[0099] Forecast Mean m The Gaussian process model is used to predict the overall system performance score corresponding to the parameter group of interest; the prediction variance. s This represents the degree of uncertainty in the prediction result. Based on the expected improvement function, we can not only find parameters with higher overall prediction performance scores, but also select unknown parameters with high uncertainty and high information gain. For each group of parameters of interest, we output the expected improvement value. The higher the expected improvement value, the greater the probability that the combination of parameters of interest has better overall performance.

[0100] In some embodiments of this application, the optimal value among all current test results is obtained, and the optimal value and a preset exploration factor are subtracted from the predicted mean to obtain the corrected improvement value; based on the corrected improvement value and the predicted variance, the expected improvement value corresponding to each group of attention parameters is determined using the expected improvement function.

[0101] Specifically, the expected improvement value EI corresponding to the parameter group of interest can be obtained through the expected improvement function according to formula (4): Formula (4): .

[0102] in, .

[0103] m This is the predicted mean output based on the Gaussian process model. s The prediction variance is based on the output of the Gaussian process model. This is the optimal value among all current test results. This is a preset exploration factor, with a value range of [0.01, 0.1]. (Z) represents the standard normal cumulative distribution function, and its value range can be [0,1]. f (Z) represents the standard normal probability density function, which can take values ​​in the range [0, 0.4]. (Z) and f (Z) are all obtained by looking up the Z value in a table.

[0104] The test set contains test results obtained from hardware performance tests on all current preset parameter groups. The optimal value selected from the test set is the [value to be determined]. .

[0105] For formula (4), if the prediction performance of a certain set of parameters of interest is... m Only slightly better The difference between the two is small, so an exploration factor is introduced. After correction, the prediction performance m Only slightly better The expected improvement value corresponding to the parameter set of focus will decrease significantly, and the algorithm will tend to select the prediction variance. s Sampling is performed on larger regions with higher uncertainty to avoid premature convergence of the optimization process to a local optimum.

[0106] As a specific implementation method, the predicted mean can be screened. m Significantly higher than The parameter group of interest, corresponding to the difference This is at a relatively high level, guiding the algorithm to perform a fine-grained search around the experimentally validated high-performance parameter range. It can also filter for prediction standard deviation. s For parameter groups with large numerical values, the measured samples in this parameter region are scarce and the model prediction information is limited, but there are undiscovered optimal parameter combinations, and the prediction variance is high. s The larger the value, the higher the probability density term. (Z) contributes more to the expected improvement value.

[0107] Returning to the embodiments of this application, the preset conditions can be set by selecting the group or the top M groups of parameters with the largest expected improvement value, or by setting all groups of parameters with expected improvement values ​​greater than the improvement threshold. This is only an example and does not constitute a limitation on the specific solution. The preset conditions can be set by oneself.

[0108] For example, the parameter group with the largest expected improvement value can be selected as the candidate parameter group.

[0109] In some embodiments of this application, after the candidate parameter group is determined, constraint verification is performed on the candidate parameter group, and hardware performance testing is performed on the candidate parameter group that passes the verification.

[0110] In other words, constraint verification can be performed by determining whether each parameter in the candidate parameter group is within its corresponding value range. During the constraint verification process, the system will record the verification results and the reasons for verification failures.

[0111] In some embodiments of this application, if the constraint verification fails, it is necessary to adjust the frequency of the candidate parameter group to make the adjusted candidate parameter group meet the test requirements, and then perform the hardware performance test on the adjusted candidate parameter group.

[0112] Specifically, for candidate parameter groups that fail verification, the frequency can be reduced by using the corresponding adjustment method based on the reason for the failure.

[0113] For example, if the failure is due to excessive temperature, PL2 and PL4 in the candidate parameter group can be reduced proportionally; if the failure is due to thermal shock risk, PL4 in the candidate parameter group can be reduced separately while keeping PL2 unchanged; if the failure is due to current constraint failure, PL2 in the candidate parameter group can be directly reduced to the current constraint upper limit.

[0114] Adjusting the frequency based on the parameters can make the adjusted candidate parameter set meet the test requirements.

[0115] This is provided as an example only and does not constitute a limitation on any specific solution.

[0116] Returning to the embodiment of this application, in step S105, hardware performance testing is performed on the candidate parameter group to determine the test results corresponding to the candidate parameter group and update them to the test set. The preset parameter group also includes the candidate parameter group.

[0117] In other words, after selecting the candidate parameter groups, the above-mentioned hardware performance tests are performed on the candidate parameter groups, and the test results corresponding to the candidate parameter groups are updated to the test set to complete the iterative update of the training test set.

[0118] For example, the total number of all preset parameter groups in the initial test set is denoted as N=2. If the number of candidate parameter groups is 1, after updating the test results corresponding to the candidate parameter groups to the initial test set, the total number of all preset parameter groups in the updated test set is N=3. If the number of candidate parameter groups is 2, after updating the test results corresponding to the candidate parameter groups to the initial test set, the total number of all preset parameter groups in the updated test set is N=4.

[0119] Meanwhile, as the test results corresponding to the candidate parameter group are updated to the test set, the optimal value in the test set may change accordingly. For example, assuming the test result is a comprehensive performance score, the optimal comprehensive performance score in the test set before the update is 0.85, and the comprehensive performance score obtained after the hardware performance test of the candidate parameter group is 0.92, then after updating the test results corresponding to the candidate parameter group to the test set, the optimal value of the updated test set will be updated to 0.92; if the comprehensive performance score of the candidate parameter group is 0.79, then the optimal value of the updated test set will remain unchanged.

[0120] In some embodiments, the total number of preset parameter groups that have been tested in the test set before the update is N, and P candidate parameter groups are obtained by filtering. After completing the hardware performance test of all P candidate parameter groups and updating the corresponding test results to the test set, the total number of preset parameter groups in the updated test set is N+P.

[0121] In step S106, the Gaussian process model is updated based on the updated test set. If the test results determined based on the current candidate parameter set meet the convergence condition, the current candidate parameter set is used as the optimization parameter set.

[0122] After each set of candidate parameter groups completes hardware performance testing, a new set of samples is added to the training test set, along with the covariance matrix. K The dimension of the Gaussian process model expands from an N-order square matrix to an N+1-order square matrix. The inverse of the covariance matrix obtained before the iteration is no longer applicable. Therefore, before each iteration, the Gaussian process model needs to be completely reconstructed based on all samples in the updated test set, and the inverse of the covariance matrix needs to be solved again. This updates the Gaussian process model. The updated Gaussian process model retains the same computational logic and can still output the predicted mean and variance for any set of parameters to be predicted. Furthermore, the prediction results incorporate the hardware performance test results corresponding to the newly added candidate parameter sets, effectively reducing the prediction uncertainty of the Gaussian process model and correcting prediction biases present before the iteration.

[0123] By continuously updating the Gaussian process model, more real hardware test data can be introduced to correct the prediction error caused by a small number of samples, continuously reduce the prediction variance in the unknown parameter region, improve the accuracy of the overall performance prediction, and accelerate the search for the optimal parameter set in the PL2 and PL4 parameter ranges.

[0124] In some embodiments, the convergence condition includes that the performance improvement magnitude corresponding to multiple consecutive iterations is less than a preset improvement threshold, wherein the performance improvement magnitude is the change in the test result determined based on the current candidate parameter set relative to the optimal value of the test set after the previous round of update.

[0125] The number of iterations and the preset improvement threshold can be configured in advance, and there are no restrictions on them.

[0126] Specifically, taking the comprehensive performance score as the test result as an example, if the comprehensive performance score corresponding to the current candidate parameter group is greater than the optimal comprehensive performance score determined in the previous round, then it continues to determine whether the performance improvement is less than the preset improvement threshold. If it is greater than the preset improvement threshold, then the convergence condition is not met, and iterative optimization continues. If it is less than the preset improvement threshold, then the number of iterations is counted once, and it is simultaneously determined whether it meets the condition of being less than the preset improvement threshold for multiple consecutive iterations. If the number of iterations is met, then it can be determined that the convergence condition is met. At this time, the current candidate parameter group is used as the optimization parameter group. If the number of iterations is not met, then it can be determined that the convergence condition is not met, and iterative optimization continues.

[0127] In a preferred embodiment, the number of iterations in the continuous multi-round iterations is 3. That is, if the performance improvement corresponding to the 3 consecutive iterations is less than the preset improvement threshold, the parameter optimization process is determined to be converged, the iterative optimization operation is terminated, and even if the iteration continues, the performance cannot be further improved.

[0128] To verify the optimization effect of the parameter optimization method in this application embodiment, the comprehensive performance score as the test result is used as an example for explanation. Specifically, the execution logic steps are as follows: Figure 2 As shown.

[0129] In step S201, the parameter value range is calculated based on the hardware operating constraints of the target computer, i.e., the value ranges of PL2 and PL4 are calculated. In step S202, a first parameter group and a second parameter group are set. In step S203, hardware performance testing is performed to construct a test set. In step S204, a Gaussian process model is constructed, and the prediction mean and prediction variance are determined using the Gaussian process model. In step S205, the expected improvement value is calculated, and a set of candidate parameter groups is selected. In step S206, constraint verification is performed on the candidate parameter groups to determine if the constraint verification passes. If not, step S207 is executed to adjust the parameter frequency. If the result of step S206 is yes, step S208 is executed to perform hardware performance testing on the candidate parameter groups and update the test set with the new test results. Step S209 continues, determining whether the comprehensive performance score obtained from the hardware performance testing of the current candidate parameter groups meets the convergence condition. If the judgment result of step S209 is yes, then step S210 is executed to output the optimized parameter group; if the judgment result of step S209 is no, then step S211 is executed to continuously update the Gaussian process model based on the updated test set, and step S204 is executed again to achieve intelligent iterative optimization.

[0130] The parameter optimization method provided in this application has higher optimization efficiency. Compared with the traditional enumeration method, it can greatly improve optimization efficiency and improve hardware operation security.

[0131] Specifically, this application generates an initial set of preset parameters based on the hardware features corresponding to different overall heat dissipation capabilities: when the first dimension of the hardware feature vector has a heat dissipation capability value of 1.35, the second dimension has a power supply capability value of 0.85, and the third dimension has a maximum power consumption logarithmic feature value of 2.08, the first set of initial parameter combinations (85W, 175W) is output through a pre-trained multivariate linear regression meta-learning model; when the first dimension of the hardware feature vector has a heat dissipation capability value of 1.15, the second dimension has a power supply capability value of 0.72, and the third dimension has a maximum power consumption logarithmic feature value of 1.95, the second set of initial parameter combinations (70W, 160W) is output through the same multivariate linear regression model.

[0132] This scheme sets the parameter combination to be predicted to include (80W, 170W), and uses an anisotropic Gaussian kernel function with differentiated length scales: PL2 dimension length scale is 5W, and PL4 dimension length scale is 10W. After iterative optimization using the parameter optimization method proposed in this application, only 15 iterations are needed, with a total time of 120 minutes, to obtain the globally optimal parameter combination (98W, 195W). Based on this optimal parameter combination, Cinebench R23 performance testing was conducted, achieving a CPU score of 20800, with a peak system temperature of only 89℃.

[0133] In contrast, the traditional full-parameter enumeration optimization scheme defines the search range based on empirical parameters (60W, 150W): PL2 selects 9 sets of values ​​in the range of 60W~100W with a step size of 5W; PL4 selects 11 sets of values ​​in the range of 150W~200W with a step size of 5W, resulting in a total of 99 combinations of parameters to be tested. The total time spent performing Cinebench R23 tests on all combinations was as long as 13.2 hours. The optimal parameter combination selected by this scheme only achieved a Cinebench R23 score of 18500, and the peak temperature of the entire machine reached as high as 94℃.

[0134] Specifically, the comparison results of the two methods are shown in Table 1: Table 1: Comparison of the effects of the parameter optimization method based on this application and the traditional enumeration method

[0135] As can be seen, the parameter optimization method provided in this application not only ensures the balance between performance and energy efficiency of parameter optimization, but also avoids the risk of hardware damage, and greatly improves optimization efficiency. It is suitable for customized optimization of CPU Turbo parameters for various laptops.

[0136] In some embodiments of this application, a parameter optimization device is provided, such as... Figure 3As shown, the parameter optimization device 300 includes a processor 301 configured to execute the steps of the parameter optimization method according to various embodiments of this application.

[0137] The processor 301 may be a processing device that includes one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor 301 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor that runs other instruction sets, or a processor that runs a combination of instruction sets. The processor 301 may also be one or more special-purpose processing devices, such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), system-on-a-chip (SoCs), etc.

[0138] In some embodiments of this application, a computer program product is provided, the computer program product comprising computer-executable instructions, which, when executed by a processor, implement the steps of the parameter optimization method according to various embodiments of this application.

[0139] This application describes various operations or functions that can be implemented as software code or instructions, or defined as software code or instructions. Such content can be directly executable source code or differential code (“incremental” or “patch” code) (“object” or “executable” form). The software code or instructions can be stored in a computer-readable storage medium and, when executed, can cause a machine to perform the described functions or operations, and include any mechanism for storing information in a machine-accessible form, such as recordable or non-recordable media (e.g., read-only memory (ROM), random access memory (RAM), disk storage media, optical storage media, flash memory devices, etc.).

[0140] The exemplary methods described in this application can be implemented, at least in part, by a machine or computer. In some embodiments, a computer-readable storage medium contains computer-executable instructions that, when executed by a processor, implement the steps of the parameter optimization methods described in various embodiments of this application.

[0141] Implementations of such methods may include software code, such as microcode, assembly language code, high-level language code, etc. Various software programming techniques can be used to create various programs or program modules. For example, program parts or program modules can be designed using or with the aid of Java, Python, C, C++, assembly language, or any known programming language. One or more of such software parts or modules can be integrated into a computer system and / or a computer-readable medium. Such software code may include computer-readable instructions for performing various methods. This software code can form part of a computer program product or a computer program module. Furthermore, in the example, the software code may be tangibly stored on one or more volatile, non-transitory, or non-volatile tangible computer-readable media, for example, during execution or at other times. Examples of such tangible computer-readable media may include, but are not limited to, hard disks, removable disks, removable optical discs (e.g., optical discs and digital video discs), magnetic tape cassettes, memory cards or memory sticks, random access memory (RAM), read-only memory (ROM), etc.

[0142] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on this application that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, which will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0143] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments can be used by those skilled in the art when reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the application. This should not be construed as an intention that a disclosed feature not claimed is necessary for any claim. Rather, the subject matter of the application may be less than all the features of a particular disclosed embodiment. Thus, the claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated as being able to be combined with each other in various combinations or arrangements. The scope of this application should be determined by reference to the appended claims and the full scope of their equivalents.

[0144] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A parameter optimization method, characterized in that, The parameter optimization method includes: Hardware performance tests are conducted based on preset parameter sets of the target computer to determine the test results corresponding to each preset parameter set in order to construct a test set. Using the preset parameter set as input and the corresponding test results as observation labels, a Gaussian process model is constructed by combining the anisotropic Gaussian kernel function; Multiple sets of parameters of interest are obtained for the target computer, and the sets of parameters of interest are input into the Gaussian process model to determine the prediction mean and prediction variance of each set of parameters of interest. Based on the predicted mean and predicted variance, the expected improvement value corresponding to each group of attention parameters is determined, and the group of attention parameters corresponding to the expected improvement value that meets the preset conditions is taken as the candidate parameter group. Hardware performance tests are performed on the candidate parameter group to determine the test results corresponding to the candidate parameter group and update them to the test set; The Gaussian process model is updated based on the updated test set. If the test results determined based on the current candidate parameter set meet the convergence condition, the current candidate parameter set is used as the optimization parameter set.

2. The parameter optimization method according to claim 1, characterized in that, The parameter optimization method further includes: determining the value range of each parameter in the parameter group based on the hardware operating constraints of the target computer.

3. The parameter optimization method according to claim 2, characterized in that, The parameter set includes short-time turbo power limit PL2 and peak turbo power limit PL4; The hardware operation constraints include current constraints, transient temperature rise constraints, and power upper limit constraints. The current constraint limits the maximum value of PL2 based on voltage and continuous current. The transient temperature rise constraint limits the instantaneous temperature rise of the hardware to no more than a threshold temperature based on the difference between PL4 and PL2. The power upper limit constraint limits PL2 to be less than PL4, and the value of PL4 is not greater than the smaller of the battery discharge power limit and the rated peak power.

4. The parameter optimization method according to claim 1, characterized in that, The preset parameter group includes a first parameter group and a second parameter group, and the parameter optimization method further includes: The hardware feature vector is determined based on the hardware operating parameters of the target computer; The first parameter set is determined based on the hardware feature vector through multiple linear regression. Select representative values ​​from the range of each parameter as the second parameter group; The hardware operating parameters include heat dissipation parameters, power supply parameters, and maximum power consumption parameters.

5. The parameter optimization method according to claim 1, characterized in that, The hardware performance test includes: After loading the corresponding parameter set on the target computer, the performance test program is run and test data is collected synchronously. Based on the test data, test results are determined to characterize the overall operating performance of the target computer.

6. The parameter optimization method according to claim 5, characterized in that, The test data includes at least the CPU sustained high load performance score, daily application scenario performance score, and average power consumption throughout the test. The parameter optimization method further includes: The CPU sustained high load performance score and the daily application scenario performance score are weighted and summed to obtain a weighted performance score. The average power consumption is weighted to obtain a weighted power consumption penalty term; The test results are determined based on the weighted performance score and the weighted power consumption penalty term.

7. The parameter optimization method according to claim 1, characterized in that, Using the preset parameter set as input and the corresponding test results as observation labels, a Gaussian process model is constructed by combining the anisotropic Gaussian kernel function, including: Different length scales are set for different parameters in the preset parameter group to obtain an anisotropic Gaussian kernel function; The anisotropic Gaussian kernel function is used to determine the similarity between each preset parameter group, and a covariance matrix is ​​constructed based on the similarity. A Gaussian process model capable of outputting predicted mean and predicted variance is obtained by training based on the covariance matrix.

8. The parameter optimization method according to claim 1, characterized in that, Based on the predicted mean and predicted variance, the expected improvement value corresponding to each group of parameters of interest is determined, including: Obtain the optimal value among all current test results, and subtract the optimal value and the preset exploration factor from the predicted mean to obtain the corrected and improved value; Based on the corrected improvement value and the predicted variance, the expected improvement value corresponding to each group of parameters of interest is determined using the expected improvement function.

9. A parameter optimization device, characterized in that, The parameter optimization apparatus includes a processor configured to perform the steps of the parameter optimization method according to any one of claims 1-8.

10. A computer program product, characterized in that, The computer program product includes computer-executable instructions that, when executed by a processor, implement the steps of the parameter optimization method according to any one of claims 1-8.

Images