WO2023082792A1

WO2023082792A1 - Parameter optimization method and apparatus

Info

Publication number: WO2023082792A1
Application number: PCT/CN2022/116726
Authority: WO
Inventors: 任宏帅; 孙涛; 杨震; 田少博; 雷晓松
Original assignee: 华为云计算技术有限公司
Priority date: 2021-11-12
Filing date: 2022-09-02
Publication date: 2023-05-19
Also published as: CN116126657A

Abstract

Embodiments of the present application provide a parameter optimization method and apparatus, the method comprising: acquiring a first candidate configuration parameter of an optimization adjustment object; running the optimization adjustment object according to the first candidate configuration parameter to obtain a first key performance indicator (KPI); acquiring a first historical performance observation record set and the variance of the first historical performance observation record set; determining a first expected value according to the first historical performance observation record set; determining a noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set; and determining a second candidate configuration parameter of the optimization adjustment object according to the first KPI, the noise variance, and the types and range of candidate configuration parameters of the optimization adjustment object. In the technical solution of the present application, a potential optimal solution may be reasonably maintained when searching for a parameter solution space, the interference of noise on the optimal solution is reduced, and parameter optimization performance is improved.

Description

Parameter optimization method and device

This application claims the priority of a Chinese patent application with application number 202111341930.X and application title "parameter optimization method and device" filed with the China Patent Office on November 12, 2021, the entire contents of which are hereby incorporated by reference in this application middle.

technical field

The present application relates to the field of computers, in particular to a parameter optimization method and device.

Background technique

In parameter optimization scenarios of software applications, middleware, algorithms, or servers, feedback from performance testing of parameters often contains noise. Especially on the cloud, the interference of noise is more serious. The real performance of the application will be affected by environmental factors such as resource interference from real physical neighbors, cabinet frequency limitation, etc., and even make the real performance of virtual resources fluctuate for a long time. Although vendors provide performance isolation and quality assurance measures, the problem of performance fluctuations is still common, and may even be exacerbated by these mechanisms.

Therefore, how to improve the parameter optimization performance is an urgent technical problem to be solved.

Contents of the invention

The present application provides a parameter optimization method and device, which can dynamically update the noise distribution during the optimization process and improve the performance of parameter optimization.

In a first aspect, a parameter optimization method is provided, the method comprising: obtaining a first candidate configuration parameter of an tuning object; running the tuning object according to the first candidate configuration parameter to obtain a first key performance indicator ( key performance indicator, KPI), the first KPI is determined according to at least one observation index; obtain the variance of the first historical performance observation record set and the first historical performance observation record set, the first historical performance observation record set includes A plurality of KPIs obtained by running the tuning object according to historical candidate configuration parameters, where the historical candidate configuration parameters are candidate configuration parameters determined before obtaining the first candidate configuration parameters; according to the first historical performance observation record set , determine the first expected value; determine the noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set; determine the noise variance according to the first KPI, the noise variance, the adjusted The type and range of the candidate configuration parameters of the optimization object are determined, and the second candidate configuration parameters of the optimization object are determined.

Tuning objects include software applications, middleware, algorithms, or servers.

The optimization algorithm determines the second candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object.

Optionally, the optimization algorithm may be Bayesian optimization algorithm, random search, grid search, cross-validation, genetic algorithm, particle swarm optimization, simulated annealing and the like.

In the embodiment of the present application, the noise variance represents the magnitude of the noise fluctuation, and the optimization algorithm can use the noise variance to consider the noise fluctuation magnitude when recommending parameters, reasonably retain the potential optimal solution, and reduce the interference of noise on the search for the optimal solution. Therefore, the second candidate configuration parameter of the tuning object is determined, and the performance of parameter optimization is improved.

With reference to the first aspect, in some implementation manners of the first aspect, before determining the first expected value according to the first historical performance observation record set, the method further includes: determining that the first KPI is not an abnormal performance observation value.

With reference to the first aspect, in some implementations of the first aspect, the method further includes: running the tuning object according to the second candidate configuration parameter to obtain a second KPI; determining that the second KPI is the The abnormal performance observation value; according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object, determine a third candidate configuration parameter of the tuning object.

Optionally, judging that the second KPI is the abnormal performance observation value may use a distance-based outlier detection method, a statistics-based outlier detection method, an unsupervised-based outlier detection method, a classification-based Cluster point detection method, outlier point detection method based on density, outlier point detection method based on information entropy, etc.

In the embodiment of the present application, by identifying and judging abnormal performance observations, identifying and filtering abnormal performance observations, avoiding the interference of abnormal performance observations on the parameter optimization process, and improving the search for optimal parameters to a certain extent efficiency.

With reference to the first aspect, in some implementation manners of the first aspect, the determining that the first KPI is not an abnormal performance observation includes: obtaining a confidence interval; determining that the first expected value is within the confidence interval; The determining that the second KPI is the abnormal performance observation value includes: determining a second expected value according to a second historical performance observation record set, wherein the second historical performance observation record set includes the first historical performance observation record A plurality of KPIs in the set and the first KPI; determine that the second expected value is outside the confidence interval.

Optionally, the obtaining the confidence interval includes: configuring the fourth candidate configuration parameter of the tuning object, and repeating the operation G times to obtain an initial performance observation record set, where G is a positive integer greater than 1; according to the initial performance observation record Ensemble to get the confidence intervals.

Optionally, the obtaining the confidence interval further includes: obtaining a preset confidence interval.

Optionally, the obtaining the confidence interval further includes: running the tuning object according to different parameter configurations to obtain a set of historical performance observation records, and constructing a confidence interval according to the set of historical performance observation records.

In the embodiment of the present application, by calculating or presetting the confidence interval, the abnormal performance observation value can be judged according to the historical performance observation records of the target software, and the abnormal performance observation value in the parameter optimization process can be filtered out, and directly determined The next group of candidate configuration parameters may select the optimal parameters to improve the efficiency of finding the optimal parameters.

With reference to the first aspect, in some implementation manners of the first aspect, the optimization is determined according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Before the second candidate configuration parameter of the object, the method further includes: determining whether the initial performance observation record set conforms to an approximate Gaussian distribution; if the initial performance observation record set does not conform to the approximate Gaussian distribution, then determining a transformation parameter, The transformation parameters can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution; the candidate configuration parameters according to the first KPI, the noise variance, and the tuning object The type and scope of determining the second candidate configuration parameters of the tuning object include: if the initial performance observation record set conforms to the approximate Gaussian distribution, then according to the first KPI, the noise variance, the the type and range of the candidate configuration parameters of the tuning object, and determine the second candidate configuration parameters of the tuning object; if the initial performance observation record set does not conform to the approximate Gaussian distribution, then use the transformation parameters to The first KPI is converted into a third KPI, and the second KPI of the tuning object is determined according to the first KPI, the third KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Candidate configuration parameters.

Optionally, the method for judging whether it is an approximate Gaussian distribution can use histogram, Shapiro-Wilk (Shapiro-Wilk, W) test, Kolmogorov-Smirnov (Kolmogorov–Smirnov, KS ) test, probability (probability-probability, P-P) diagram, quantile-quantile (quantile-quantile, Q-Q) diagram, etc.

It should be understood that a degree threshold can be preset, and when the value obtained by using the method for judging an approximate Gaussian distribution is less than the degree threshold, it is a non-Gaussian distribution, and when it is greater than or equal to the degree threshold, it is an approximate Gaussian distribution. Exemplarily, when the KS test is used, when the error probability (P) value of accepting the null hypothesis is greater than or equal to 0.05, it is judged that the distribution conforms to an approximate Gaussian distribution.

Optionally, the method of transformation may include beta (beta) transformation, logarithm (log) transformation, square root transformation, taking <1 to a certain power transformation, reciprocal transformation, square root anti-rotation transformation, Box-Cox ( Box-Cox) transform, inverse transform sampling, inverse probability integral transform and Fisher transform, etc.

With reference to the first aspect, in some implementation manners of the first aspect, the optimization is determined according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Before the third candidate configuration parameter of the object, the method further includes: determining whether the initial performance observation record set conforms to an approximate Gaussian distribution; if the initial performance observation record set does not conform to the approximate Gaussian distribution, then determining a transformation parameter, The transformation parameters can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution; the candidate configuration parameters according to the first KPI, the noise variance, and the tuning object The type and scope of determining the third candidate configuration parameters of the tuning object include: if the initial performance observation record set conforms to the approximate Gaussian distribution, then according to the first KPI, the noise variance, the The type and range of the candidate configuration parameters of the tuning object, determine the third candidate configuration parameters of the tuning object; if the initial performance observation record set does not conform to the approximate Gaussian distribution, use the transformation parameters to convert the The first KPI is converted into a fourth KPI, and the third KPI of the tuning object is determined according to the first KPI, the fourth KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Candidate configuration parameters.

In the embodiment of the present application, the impact of noise can be dealt with during the parameter optimization process, and targeted processing can be performed according to the characteristics of the noise, by converting the non-Gaussian distribution of performance observation records into approximate Gaussian distribution of performance observation records , to suppress the impact of non-Gaussian distribution noise on parameter optimization, and can deal with time-varying noise, which improves the efficiency of finding optimal parameters to a certain extent.

With reference to the first aspect, in some implementation manners of the first aspect, the determining the noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set includes: according to The following formula determines the noise variance:

Wherein, i is the serial number of the candidate configuration parameter of the tuning object,

is the variance of the first historical performance observation record set, E(xi ₎ is the expected value of the first historical performance observation record set, y _i is the first KPI,

is the noise variance, and y _i -E(xi ₎ is the residual error between the expected value of the first historical performance observation record set and the first KPI.

Optionally, the formula for determining the noise variance can also be

or

Among them, P is the penalty coefficient.

In the embodiment of the present application, the noise distribution can be dynamically estimated, and the noise variance of this time can be obtained through the expected value of the first historical performance observation record set, the residual of the first KPI, and the variance of the first historical performance observation record set. The optimization algorithm can use the noise variance to consider the noise fluctuation range when searching the parameter solution space, reasonably retain the potential optimal solution, and reduce the interference of noise on the search for the optimal solution, thereby improving the parameter optimization performance.

In a second aspect, an embodiment of the present application provides a computer device, where the computer device includes a unit for implementing the first aspect or any possible implementation manner of the first aspect.

In a third aspect, an embodiment of the present application provides a computer device, the computer device includes a processor, the processor is used to be coupled with a memory, read and execute instructions and/or program codes in the memory, to perform the first aspect or Any possible implementation of the first aspect.

In a fourth aspect, an embodiment of the present application provides a chip system, the chip system includes a logic circuit, the logic circuit is used to couple with an input/output interface, and transmit data through the input/output interface to perform the first aspect or the first Any possible implementation of the aspect.

In the fifth aspect, the embodiment of the present application provides a computer-readable storage medium, the computer-readable storage medium stores program codes, and when the computer storage medium is run on a computer, the computer executes the first aspect or the first aspect. any possible implementation of .

In a sixth aspect, an embodiment of the present application provides a computer program product, the computer program product comprising: computer program code, when the computer program code is run on a computer, the computer is made to execute any of the first aspect or the first aspect. One possible implementation.

Description of drawings

FIG. 1 is an example diagram of an application of a parameter optimization method provided in an embodiment of the present application.

Fig. 2 is an exemplary flow chart of a parameter optimization method provided by an embodiment of the present application.

Fig. 3 is a schematic diagram of setting parameter types and ranges provided by the embodiment of the present application.

Fig. 4 is a schematic diagram of the running result of the tuning object provided by the embodiment of the present application.

Fig. 5 is an exemplary flow chart of the internal processing of the noise diagnosis module during the first iteration provided by the embodiment of the present application.

Fig. 6 is an example histogram of noise distribution estimation provided by the embodiment of the present application.

Fig. 7 is an example histogram conforming to the approximate Gaussian distribution provided by the embodiment of the present application.

Fig. 8 is an example histogram of the non-Gaussian distribution provided by the embodiment of the present application.

Fig. 9 is an exemplary flow chart of the internal processing of the noise diagnosis module during the second iteration provided by the embodiment of the present application.

Fig. 10 is an exemplary flow chart of the internal processing of the noise diagnosis module at the thirtieth iteration provided by the embodiment of the present application.

Fig. 11 is an exemplary flow chart of the internal processing of the noise diagnosis module at the sixtieth iteration provided by the embodiment of the present application.

FIG. 12 is a structural example diagram of a computer device provided by an embodiment of the present application.

Fig. 13 is a structural example diagram of another computer device provided by an embodiment of the present application.

Fig. 14 is an exemplary diagram of a computer program product provided by an embodiment of the present application.

Detailed ways

The final performance observations of a software application, middleware, algorithm, or server may be determined using the mean, mode, or median of performance observations obtained from multiple validations. However, the types of noise are not static, and only one estimation method cannot be used for all applications.

The above method has the following problems:

Calculating the mean, mode, and median in a statistical sense requires a certain amount of data, and the real application software parameter verification often has fewer verification times due to cost reasons, and a small number of verification times cannot correctly estimate these statistical measures. The distribution of noise may change over time, and establishing a fixed noise model for noise cannot correctly weaken the influence of noise. Different types of noise affect the tuning algorithm in different ways and require different methods to deal with it.

For example: in the scenario of GCC compiler optimization, due to the long running time of the program code, the running time of each type of program is affected by the optimization switch item, ranging from more than ten minutes to several hours, this situation cannot be used More validation times to estimate the mean and median.

In the middleware tuning scenario, take the tuning of middleware kafka or tomcat as an example. When kafka or tomcat is affected by virtualized resources on the cloud, competition for virtualized resources on the cloud will occur during the peak period of request processing, resulting in The performance is degraded, making the performance of the application fluctuate. This situation is due to changes in performance caused by external factors. Traditional methods cannot distinguish between real performance changes and performance changes due to noise interference.

The traditional method of statistically calculating the average value of multiple verifications can be applied to scenarios with fewer verifications. However, if encountering abnormal noise that seriously deviates from the distribution range of performance observations, it will cause the average value of statistical performance observations to deviate from the true value.

Statistical methods can deal with the impact of abnormal noise by counting the mode or median through multiple verifications, but this method requires more verifications and the cost of verification will increase accordingly.

For this reason, this application proposes a parameter optimization method, which can dynamically update the noise distribution during the tuning process and improve the parameter optimization performance.

The technical solution provided by this application can be applied to parameter tuning scenarios affected by noise data in the computer field, including middleware parameter tuning, software parameter tuning, and server parameter tuning, etc., especially for parameter tuning on the cloud Scenes.

In order to solve various types of noise more specifically, this application divides noise into two types: abnormal noise and normal noise according to the characteristics of noise.

Anomaly noise is the heavily counterfactual outlier observations produced by an anomaly problem. Abnormal noise is usually sporadic noise caused by accidental factors such as program abnormality or hardware failure. The frequency of abnormal noise is low but the amplitude is large, and a small amount of repeated verification is difficult to deal with. The mean size of abnormal noise is disturbed by large noise.

Problems caused by abnormal noise: In the absence of sufficient data support, it is difficult to judge whether the performance jitter is caused by parameter effects or abnormal noise; when the problem model is complex, due to the small number of interactive verifications, the method of estimating performance and calculating confidence intervals will fail.

Normal noise: Due to the virtualization of resources on the cloud and complex environmental impacts, the real performance of the monitored target software has been fluctuating. This kind of noise cannot be eliminated for a long time, and it is affected by many variables, so the observed value of the feedback cannot correctly reflect the real performance of the parameters. Normal noise accompanies for a long time, the noise amplitude is small, and the mean method can be approximated, but the verification cost is high.

Normal noise can be divided into three types:

a) Gaussian noise: This kind of noise exists for a long time and conforms to a fixed Gaussian distribution. The traditional parameter optimization algorithm takes Gaussian noise as the premise assumption, and believes that the observed value sampled each time has a high probability of being close to the real value.

Problems caused by Gaussian noise: Although the premise assumption of Gaussian noise can statistically guarantee that the sampling point has a high probability of falling in the mode histogram interval, there is still the possibility of inaccurate sampling; although the noise conforms to the Gaussian distribution, the fluctuation interval of the noise is still Very broad, a small number of repeated verifications cannot correctly estimate its true performance.

b) Non-Gaussian noise: Due to the interference of multi-party noise, the noise distribution does not follow a specific variance distribution, and even multiple peaks appear during statistics. When using mean statistics, a large amount of verification data support is required.

Problems caused by non-Gaussian noise: noise does not follow a specific distribution, it is difficult to model noise; a large number of repeated verifications are required to statistically distribute histograms.

c) Timing change noise: Since the use of certain services in the cloud environment is affected by user access patterns, the use of resources presents a regular pattern of timing. Therefore, during the tuning test, these peak resource preemption may cause The observation noise model changes over time.

Problems caused by timing variation noise: It is difficult to model the noise that changes with time, and it is difficult to locate the key factors that cause timing variation; the noise distribution may change, and it is impossible to guarantee that the noise will always maintain the same change law.

The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings. Apparently, the described embodiments are some of the embodiments of the present application, not all of them. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the scope of protection of this application.

The application example of the parameter optimization method shown in FIG. 1 includes an tuning server 100 and a client 101 , wherein the tuning server 100 includes an tuning service module 110 , a noise diagnosis module 140 and a tuning record storage module 150 . The client 101 includes a performance monitoring module 130 and a tuning object 120 .

The tuning service module 110 can recommend candidate configuration parameters by receiving data from the noise diagnosis module 140 , and save the data received from the noise diagnosis module 140 in the tuning record storage module 150 .

The tuning record storage module 150 may be a hard drive or other computer-readable storage medium. The tuning record storage module 150 may store data sent by the tuning service module 110, or information such as instructions.

The tuning object 120 may be a middleware, a software application, an algorithm, or a server that needs parameter tuning, which is not limited in this application. The tuning object 120 can run the application according to the candidate configuration parameters recommended by the tuning service module 110 .

The performance monitoring module 130 may be a tool with a monitoring function, an application, or a script code capable of monitoring performance. The performance monitoring module 130 can monitor the corresponding KPIs generated when the tuning object 120 is running according to the KPIs preset by the user, and transmit the obtained KPIs to the noise diagnosis module 140 .

The noise diagnosis module 140 can perform noise diagnosis and noise transformation on the KPI according to the historical performance observation record obtained from the tuning record storage module 150 and the KPI sent by the performance monitoring module 130, and send the obtained denoised KPI and noise variance to to the tuning service module 110.

Next, the present application will describe in detail the interaction and work flow of each module on the basis of the application example shown in FIG. 1 .

210. Define elements and a preset number of iterations M, where M is a positive integer greater than 1.

It should be understood that defining elements may include configuring parameters, setting KPI targets and service loads, and the like.

The configuration parameter can be to configure the tuning parameter type and its range on the tuning script. As an example, Figure 3 is a schematic diagram of setting parameter types and ranges for kafka parameter tuning provided by the embodiment of this application. The parameter types may include integer, floating point, Boolean, enumeration, etc. The range of can be set according to the type of parameter. For example, when the parameter log.cleaner.enable is set to Boolean, the parameter log.cleaner.enable is true or false. The value of the parameter is within a preset range. For example, when the range of the parameter message.max.bytes is set to 1K-1M, the value of the parameter message.max.bytes is between 1K and 1M.

The KPI in setting the KPI target refers to the KPI index for parameter optimization. The KPI index can be one observation index or a combination of several observation indexes. As an example, KPI indicators can be specified as commonly used performance test indicators, including producer throughput or throughput rate, transaction average response time, transaction success rate, CPU utilization, maximum number of jobs (max-jobs), and number of key jobs ( critical-jobs) and so on. Users can set KPI indicators according to requirements, which is not limited in this application. Exemplarily, when the KPI index is set as the producer throughput (transactions per second, TPS), the optimal parameter found in the embodiment of the present application is the parameter corresponding to the maximum TPS. When setting the KPI indicator as a weighted combination of multiple indicators, these indicators need to have the same dimension. Exemplarily, when tuning the Java Virtual Machine (Java Virtual Machine, JVM) public basic test suite Specjbb, the KPI indicator can be the combination of these two indicators max-jobs and critical-jobs, and the KPI can be set to KPI=max- jobs+critical-jobs. Different weights can also be set for these indicators, for example, KPI=max-jobs*0.8+critical-jobs*0.2. The optimal parameter to be found at this time is the parameter corresponding to when the KPI is the largest.

Optionally, the business load of the tuning object 120 may also be set, and a stress test or a load test may be performed on the tuning object 120 or a code program specified by the tuning object 120 . Stress test is to obtain the maximum service level test that the system can provide by determining the bottleneck or unacceptable performance point of a system. Load test is to verify that the test object is under different operating conditions (such as different Acceptability of performance behavior under the number of users, number of transactions, etc.). Optimize parameters based on performance test results obtained from stress tests or load tests.

Optionally, the user can preset the number of iterations M of the optimization method according to requirements, or use a default value of the number of iterations. The number of iterations is used to limit the number of rounds the optimization method learns. It can be understood that the greater the number of iterations M, the more likely it is to find optimal parameters, but the energy consumption may also increase accordingly. Therefore, users can set the number of iterations M according to their actual needs. Exemplarily, M can be set to 30, 50, 60 and so on.

220. The tuning service module 110 recommends a group of candidate configuration parameters.

Optionally, the tuning service module 110 provides at least one optimization algorithm, and the optimization algorithm may include Bayesian optimization algorithm, random search, grid search, cross-validation, genetic algorithm, particle swarm optimization, simulated annealing, etc. Applications are not limited to this. When the number of learning iterations of the current parameter optimization method is less than M, the optimization algorithm in the tuning service module 110 can recommend the next set of candidate configuration parameters for the tuning object 120 to run according to the input provided by the noise diagnosis module.

It should be understood that at least one parameter is included in a set of candidate configuration parameters. Exemplarily, the tuning service module 110 can execute the optimization algorithm once to obtain a set of candidate configuration parameters for recommendation, or execute the optimization algorithm multiple times to obtain multiple sets of candidate configuration parameters and select one of them for recommendation. The selection method can be The selection is made randomly or according to the time when the candidate configuration parameters are obtained, which is not limited in this application.

230. The tuning object 120 configures the parameter and performs a test.

The tuning object 120 of the client obtains and configures a set of candidate configuration parameters recommended by the tuning service module 110 , and then runs the configuration to obtain the KPI corresponding to the set of parameters. Exemplarily, when the KPI indicator is the throughput of the producer, the KPI corresponding to this group of parameters may be 130M/s, 158M/s, etc. at this time. It should be understood that the KPIs given here are only examples, and the specific values are subject to the results of actual operation.

240 , the performance monitoring module 130 sends the KPI to the noise diagnosis module 140 .

The performance monitoring module 130 monitors the tuning object 120 according to a preset KPI index. Exemplarily, when the KPI indicator is the throughput of the producer, monitor the TPS when the tuning object 120 is running, and then send the TPS value to the noise diagnosis module 140 .

250. The noise diagnosis module 140 performs diagnosis transformation on the KPI.

The noise diagnosis module 140 can deal with the impact of noise during the process of parameter optimization. Targeted processing is carried out according to the characteristics of different noises. If the KPI obtained in this operation is judged to be abnormal noise, the abnormal performance observation value is filtered, and the tuning service module 110 directly recommends the next set of candidate configuration parameters. If it is judged that the KPI obtained in this operation is normal noise, the normal noise is suppressed according to the requirements and its expected value and noise variance are calculated, and the denoised KPI and noise variance are output.

260. The tuning service module 110 saves the current tuning result.

The tuning service module 110 saves the parameters configured by the tuning object 120 this time and the denoised KPI sent by the noise diagnosis module 140 in the tuning record storage module 150 .

270. Determine whether the number of iterations is less than M.

The tuning service module 110 judges whether the number of iterations this time is less than the number of iterations preset by the user M, if the number of iterations this time is less than M, then use the optimization algorithm to recommend the next group of candidate configurations according to the denoised KPI and noise variance parameter. After the optimization algorithm obtains the noise variance, the error interval can be considered in the optimization process, a certain tolerance to the local optimal point can be maintained, and the potential parameter space that may be close to the optimal point can be reserved.

If the number of iterations this time is greater than or equal to the number of iterations M preset by the user, go to step 280 .

280, select the optimal parameters.

If the number of iterations at this time is greater than or equal to the number of iterations preset by the user M, it means that the iterative learning process of the optimization method has met the needs of the user. At this time, the set of candidate configuration parameters corresponding to the optimal KPI in the M round output as a choice of optimal parameters.

When the number of learning iterations of the current parameter optimization method is greater than or equal to M, the tuning service module 110 selects a set of configuration parameters used when the performance of the tuning object 120 is optimal from the tuning record storage module 150, the configuration parameters is the optimal parameter. Exemplarily, when the KPI indicator is set to TPS, the optimal parameter found in this embodiment of the present application is the parameter corresponding to the maximum TPS.

Next, on the basis of the application example shown in FIG. 1 , the present application assumes that the number of iterations preset by the user is M=60, and m is the current number of iterations. Taking m=1, m=2, m=30 and m=60 as examples, the internal processing flow of the noise diagnosis module is described in detail.

It should be understood that when m=i, the candidate configuration parameter used by the tuning object 120 may be referred to as a candidate configuration parameter i, where i is a positive integer greater than or equal to 1. Exemplarily, when m=1, the candidate configuration parameter used by the tuning object 120 may be called candidate configuration parameter 1, when m=2, the candidate configuration parameter used may be called candidate configuration parameter 2, and when m=30, the candidate configuration parameter used The parameters may be referred to as candidate configuration parameters 30 , and the candidate configuration parameters for m=60 may be referred to as candidate configuration parameters 60 .

Optionally, in the initial stage when the parameter optimization method has not started iterative learning, that is, before the tuning object 120 runs according to the candidate configuration parameter 1, the same set of configuration parameters can be used to make the tuning object 120 repeatedly run multiple times, the set The configuration parameters may be referred to as default configuration parameters, and the default configuration parameters may be parameters set by users. The number of repetitions can be set by the user, or the default number of repetitions can be used. Exemplarily, the number of repeated operations may be set to 30 times, 50 times, 80 times, etc., which is not limited in this application. The initial performance observation record set corresponding to the default parameter configuration can be obtained through the repeated operations.

Exemplarily, FIG. 4 is a schematic diagram of the result of repeated operation of the tuning object 120 provided by the embodiment of the present application according to the configuration of the same set of parameters. The set of default configuration parameters is set to "num.io.threads=12, num.network.threads= 6, socket.receive.buffer.bytes=102400". All the corresponding KPI values in the figure are the initial performance observation records set corresponding to the set of default configuration parameters.

The goal of the parameter optimization method is to find the parameters corresponding to the optimal KPI, as shown in formula (1),

x ^* ＝argmax _i∈N (f(x _i )+∈ _i ) ∈～N(0,σ ² ) (1)

x ^* is the candidate configuration parameter corresponding to the optimal KPI. Wherein i is the serial number of the candidate configuration parameter used by the current tuning object 120, i is equal to the current iteration number m, x is the configuration parameter, x _i is the candidate configuration parameter i of the tuning object 120, and f is the model of the tuning object 120 , f( _xi ) is the real performance value of the tuning object 120 running according to the candidate configuration parameter _xi , and ∈ _i is the noise value of the tuning object 120 running according to the candidate configuration parameter _xi . As shown in formula (2), the sum of the true performance value f(xi ₎ and its corresponding noise value ∈ _i is the KPI value y _i .

y _i =f(x _i )+∈ _i (2)

It should be understood that, as shown by formulas (3) to (5), the initial performance observation record set Y is a set of all KPIs obtained by using the same set of default configuration parameters X to run repeatedly for a specified number of times. Variance is the variance, and _ci is a constant. Since the real performance value f( _xi ) obtained by running the same set of parameter configurations is constant, the variance of the initial performance observation record set Y is equal to the noise variance corresponding to the default configuration parameters _xi of the tuning object 120 running. initial noise variance

It is equal to the variance of the initial performance observation record set Y.

variance(Y)＝variance(f(x _i )+∈ _i ) ∵f(x _i )＝c _i (3)

Fig. 5 is an exemplary flow chart of the internal processing of the noise diagnosis module provided by the embodiment of the present application when m=1. When m=1, the tuning object 120 can run according to the candidate configuration parameter 1 . The performance monitoring module 130 monitors the tuning object 120 according to the preset KPI index, obtains the KPI corresponding to the candidate configuration parameter 1, and sends the KPI corresponding to the candidate configuration parameter 1 to the noise diagnosis module. The noise diagnosis module 140 may execute the process shown in FIG. 5 according to the received KPI.

510. Estimating the noise distribution.

Optionally, 510 noise distribution estimation may include two steps, judging whether the initial performance observation record set conforms to an approximate Gaussian distribution and calculating the initial noise variance of the initial performance observation record set

Judging whether the initial performance observation record set conforms to an approximate Gaussian distribution: Estimate the noise distribution of the initial performance observation record set obtained by repeated operations of the tuning object 120 and determine whether it is an approximate Gaussian distribution. The judgment method used in the embodiment of the present application is a histogram. Exemplarily, FIG. 6 is an example histogram of noise distribution estimation provided by the embodiment of the present application. Each tuning test can have only one KPI, and the obtained histogram is the histogram corresponding to the KPI. KPI can be set as a single observation index or a combination of multiple observation indexes. When set as a combination of multiple observations, the multiple observations need to have the same dimension. That is to say, one KPI can correspond to one observation indicator, or can correspond to multiple observation indicators.

Exemplarily, when the KPI is set as a single observation index and is the throughput of the producer, the KPI at this time corresponds to an observation index 1, and the observation index 1 is the throughput of the producer. Schematically, FIG. 7 is a throughput histogram conforming to an approximate Gaussian distribution obtained by running the tuning object 120 .

Exemplarily, when the KPI indicator is set as a combination of n observation indicators, then the KPI has n corresponding observation indicators. For example, when the KPI index is set as KPI=max-jobs+critical-jobs, the KPI at this time corresponds to two observation indexes, observation index 1 is the maximum number of jobs, and observation index 2 is the number of critical jobs. Schematically, FIG. 8 is a KPI histogram of a non-Gaussian distribution obtained by running the tuning object 120 .

Optionally, the method for judging whether it is an approximate Gaussian distribution can also use W test, KS test, P-P diagram, Q-Q diagram, etc., which is not limited in this application.

If the initial performance observation record set is non-Gaussian distribution, the transformation parameters are determined. The transformation parameter can transform the initial performance observation record set into an initial performance observation record set 2 conforming to an approximate Gaussian distribution. If the initial performance observation record set conforms to an approximate Gaussian distribution, then the transformation parameter does not need to be determined.

Optionally, the transformation parameter may be determined by transforming the initial performance observation record set. For example, different parameters may be used to transform the initial performance observation record set. If the performance observation record set transformed according to a certain parameter conforms to an approximate Gaussian distribution, then the parameter may be determined as the transformation parameter. The transformation method from non-Gaussian distribution to approximate Gaussian distribution can be any of the following transformations: beta transformation, log transformation, square root transformation, power transformation of a certain number <1, reciprocal transformation, square root anti-rotation transformation, Box-Cox transformation, Inverse transformation sampling, inverse probability integral transformation, Fisher transformation, etc., which method is used for transformation is not limited in this application.

Optionally, the step of judging whether the initial performance observation record set conforms to an approximate Gaussian distribution may also be performed after 520 .

Computes the initial noise variance for the initial set of performance observation records

And used for subsequent noise diagnosis.

520. Outlier detection.

Optionally, the initial noise variance obtained in 510 noise distribution estimation can be used

Constructs the confidence interval for the initial set of performance observation records according to the confidence level set by the user. Gaussian fitting is then performed on the set of historical performance observation records of the tuning object 120 . The set of historical performance observation records includes all KPIs obtained from the historical operation of the tuning object 120 except the KPI obtained according to the current candidate configuration parameter operation. It should be understood that the historical performance observation record set when m=1 is the same as the initial performance observation record set.

Optionally, the method of performing Gaussian fitting on the set of historical performance observation records includes performing function approximation on function points, decision tree method, Gaussian curve fitting and other methods with fitting functions, or using some software that can be used to fit data interface, which is not limited in this application.

After performing Gaussian fitting on the historical performance observation record set, the expected value of candidate configuration parameter 1 can be obtained, and then judge whether the expected value falls within the confidence interval. If it falls within the confidence interval, the KPI corresponding to candidate configuration parameter 1 is not abnormal noise. If If it falls outside the confidence interval, the KPI corresponding to candidate configuration parameter 1 is abnormal noise, and abnormal noise is screened out in this way.

Optionally, the method for filtering out abnormal noise can also use a distance-based outlier detection method, a statistics-based outlier detection method, an unsupervised-based outlier detection method, a classification-based cluster detection method, a method based on The outlier detection method based on density, the outlier detection method based on information entropy, etc., are not limited in this application.

It should be understood that if the KPI corresponding to the candidate configuration parameter 1 is judged to be abnormal noise, the tuning service module 110 directly recommends the next group of candidate configuration parameters 2 .

The above-mentioned technical solution can deal with the impact of noise in the process of parameter optimization, and carry out targeted processing according to the characteristics of noise, identify and filter abnormal noise, and avoid the interference of abnormal noise on the parameter optimization process, to a certain extent Improve the efficiency of finding the optimal parameters.

If the KPI corresponding to candidate configuration parameter 1 is judged not to be abnormal noise, then follow-up steps can be performed according to whether the initial performance observation record set conforms to an approximate Gaussian distribution.

For example, if the initial performance observation record set does not conform to the approximate Gaussian distribution, then step 530 and step 540 may be performed; if the initial performance observation record set conforms to the approximate Gaussian distribution, then step 540 may be directly performed.

530. Use the transformation parameters determined in step 510 to transform the KPI corresponding to the candidate configuration parameter 1.

The above technical solution can deal with the impact of noise in the process of parameter optimization, and according to the characteristics of the noise, it can be processed in a targeted manner, by converting the KPI of non-Gaussian distribution into the value of approximate Gaussian distribution, and suppressing the impact of non-Gaussian distribution noise on the parameters. The impact of optimization, being able to deal with time-varying noise, improves the efficiency of finding optimal parameters to a certain extent.

540. Calculate noise variance when m=1.

As shown by formulas (6) to (8), in the interactive tuning process of dynamic noise estimation, Gaussian fitting can be performed on the historical performance observation record set, and the expected value E(xi ₎ can be estimated for the current candidate configuration parameters x _i At this time, the KPI value y _i and the expected value are used to calculate the residual, and the new residual result is used to update the noise variance

Among them, i is the current iteration number m,

The noise variance of the last iteration computed for any of equations (6) to (8).

The method of calculating the expected value E( _xi ) can be regression models such as GP Gaussian process regression, linear regression, polynomial regression, ridge regression, least absolute shrinkage and selection operator (Lasso) regression, or regression tree model, limit Random forest methods such as extreme gradient boosting (XGBoost), categorical boosting (CatBoost), and light gradient boosting machine (LightGBM) are not limited in this application.

P in formula (8) is a default or user-set penalty coefficient. Exemplarily, the penalty coefficient may be an integer greater than 0 or a floating point type, which is not limited in the present application.

It should be understood that formulas (6) to (8) are only examples of formulas for updating noise variance, and those of ordinary skill in the art may use variations of similar formulas or different formulas to realize the described functions in combination with the descriptions of the embodiments disclosed herein , but such implementation should not be considered beyond the scope of this application.

When m=1, i=1. Exemplarily, when formula (6) is used to update the noise variance, E(x ₁ ) may be an expected value calculated by using GP Gaussian process regression to perform Gaussian fitting on the initial performance observation record set. y ₁ is the KPI corresponding to candidate configuration parameter 1, and y ₁ -E(x ₁ ) is the residual difference between the KPI corresponding to candidate configuration parameter 1 and the expected value.

is the initial noise variance,

is the variance corresponding to candidate configuration parameter 1, that is, the noise variance output by the noise diagnosis module 140 this time.

After the noise diagnosis module, the KPI value y ₁ and the noise variance will be output

to the tuning service module 110. The optimization algorithm in the tuning service module 110 can pass the KPI value y ₁ , and the noise variance

When recommending the next set of candidate configuration parameters 2, consider the magnitude of noise fluctuations, reasonably retain the potential optimal solution, and reduce the interference of noise on finding the optimal solution, thereby improving the performance of parameter optimization.

If the KPI value y ₁ has undergone transformation from a non-Gaussian distribution to an approximate Gaussian distribution in 530 , the noise diagnosis module will also output the transformed KPI value y ₁ to the tuning service module 110 . The optimization algorithm in the tuning service module 110 can pass the KPI value y ₁ , the transformed KPI value y ₁ and the noise variance

Fig. 9 is an exemplary flow chart of the internal processing of the noise diagnosis module provided by the embodiment of the present application when m=2. When m=2, the tuning object 120 may operate according to the candidate configuration parameter 2 . The performance monitoring module 130 monitors the tuning object 120 according to the preset KPI index, obtains the KPI corresponding to the candidate configuration parameter 2, and sends the KPI corresponding to the candidate configuration parameter 2 to the noise diagnosis module. The noise diagnosis module 140 may execute the process shown in FIG. 9 according to the received KPI.

710. Outlier detection.

It should be understood that the historical performance observation record set when m=2 includes the initial performance observation record set and the KPI corresponding to candidate configuration parameter 1. If the KPI corresponding to candidate configuration parameter 1 is abnormal noise, then the historical performance observation record set when m=2 only includes the initial performance observation record set, and Gaussian fitting is performed on the historical performance observation record set to obtain the expected value of candidate configuration parameter 2, Then, according to the confidence interval constructed in step 520 in Figure 5, it is judged whether the expected value falls within the confidence interval. If it falls within the confidence interval, the KPI corresponding to candidate configuration parameter 2 is not abnormal noise. If it falls outside the confidence interval, the candidate configuration parameter The KPI corresponding to 2 belongs to abnormal noise, and the abnormal noise is screened out in this way.

It should be understood that if the KPI corresponding to the candidate configuration parameter 2 is judged to be abnormal noise, the tuning service module 110 directly recommends the next group of candidate configuration parameters 3 .

If the KPI corresponding to the candidate configuration parameter 2 is judged not to be abnormal noise, then follow-up steps can be performed according to whether the initial performance observation record set conforms to an approximate Gaussian distribution.

For example, if the initial performance observation record set does not conform to the approximate Gaussian distribution, then step 720 and step 730 may be performed; if the initial performance observation record set conforms to the approximate Gaussian distribution, then step 730 may be directly performed.

720. Use the transformation parameters determined in step 510 to transform the KPI corresponding to the candidate configuration parameter 2.

730. Calculate noise variance when m=2.

When m=2, i=2. Exemplarily, when formula (6) is used to update the noise variance, E(x ₂ ) may be an expected value calculated by using GP Gaussian process regression to perform Gaussian fitting on the set of historical performance observation records when m=2. y ₂ is the KPI corresponding to the candidate configuration parameter 2, and y ₂ -E(x ₂ ) is the residual difference between the KPI corresponding to the candidate configuration parameter 2 and the expected value.

is the noise variance corresponding to candidate configuration parameter 1,

The variance corresponding to candidate configuration parameter 2, that is, the noise variance of this output.

Optionally, the method for calculating the expected value E(x ₂ ) can also be regression models such as linear regression, polynomial regression, ridge regression, and Lasso regression, or regression tree models, random forest methods such as XGBoost, CatBoost, and LightGBM. Not limited.

After the noise diagnosis module, the KPI value y ₂ and the noise variance will be output

to the tuning service module 110. The optimization algorithm in the tuning service module 110 can pass the KPI value y ₂ , and the noise variance

When recommending the next set of candidate configuration parameters 3, the amplitude of noise fluctuations is considered, and the potential optimal solution is reasonably reserved to reduce the interference of noise on finding the optimal solution, thereby improving the performance of parameter optimization.

If the KPI value y ₂ has been transformed from a non-Gaussian distribution to an approximate Gaussian distribution in 720 , the noise diagnosis module will also output the transformed KPI value y ₂ to the tuning service module 110 . The optimization algorithm in the tuning service module 110 can pass the KPI value y ₂ , the transformed KPI value y ₂ and the noise variance

Fig. 10 is an exemplary flow chart of the internal processing of the noise diagnosis module provided by the embodiment of the present application when m=30. When m=30, the tuning object 120 can operate according to the candidate configuration parameters 30 . The performance monitoring module 130 monitors the tuning object 120 according to the preset KPI index, obtains the KPI corresponding to the candidate configuration parameter 30, and sends the KPI corresponding to the candidate configuration parameter 30 to the noise diagnosis module. The noise diagnosis module 140 may execute the process shown in FIG. 10 according to the received KPI.

810. Outlier detection.

It should be understood that the historical performance observation record set when m=30 includes the initial performance observation record set and KPIs corresponding to candidate configuration parameters 1 to 29 after removing abnormal noise. Perform Gaussian fitting on the set of historical performance observation records to obtain the expected value of the candidate configuration parameter 30, and then judge whether the expected value falls within the confidence interval based on the confidence interval constructed in step 520 in Figure 5, and if it falls within the confidence interval, then the candidate configuration The KPI corresponding to the parameter 30 is not abnormal noise. If it falls outside the confidence interval, the KPI corresponding to the candidate configuration parameter 30 is abnormal noise. The abnormal noise is screened out in this way.

It should be understood that if the KPI corresponding to the candidate configuration parameters 30 is judged to be abnormal noise, the tuning service module 110 directly recommends the next group of candidate configuration parameters 31 .

If the KPI corresponding to the candidate configuration parameter 30 is judged not to be abnormal noise, the subsequent steps may be performed according to whether the initial performance observation record set conforms to an approximate Gaussian distribution.

For example, if the initial performance observation record set does not conform to the approximate Gaussian distribution, then step 820 and step 830 may be performed; if the initial performance observation record set conforms to the approximate Gaussian distribution, then step 830 may be directly performed.

820. Use the transformation parameters determined in step 510 to transform the KPIs corresponding to the candidate configuration parameters 30.

830. Calculate the noise variance when m=30.

When m=30, i=30. Exemplarily, when formula (6) is used to update the noise variance, E(x ₃₀ ) may be an expected value calculated by using GP Gaussian process regression to perform Gaussian fitting on the set of historical performance observation records when m=30. y ₃₀ is the KPI corresponding to the candidate configuration parameter 30, and y ₃₀ -E(x ₃₀ ) is the residual difference between the KPI corresponding to the candidate configuration parameter 30 and the expected value.

is the noise variance corresponding to the candidate configuration parameter 29,

is the variance corresponding to the candidate configuration parameter 30, that is, the noise variance output this time.

Optionally, the method for calculating the expected value E(x ₃₀ ) can also be regression models such as linear regression, polynomial regression, ridge regression, and Lasso regression, or random forest methods such as regression tree models, XGBoost, CatBoost, and LightGBM. Not limited.

After the noise diagnosis module, the KPI value y ₃₀ and the noise variance will be output

to the tuning service module 110. The optimization algorithm in the tuning service module 110 can pass the KPI value y ₃₀ , and the noise variance

When recommending the next set of candidate configuration parameters 31, consider the magnitude of noise fluctuations, reasonably retain the potential optimal solution, reduce the interference of noise on finding the optimal solution, and improve the performance of parameter optimization.

If the KPI value y ₃₀ has undergone transformation from a non-Gaussian distribution to an approximate Gaussian distribution in 820 , the noise diagnosis module will also output the transformed KPI value y ₃₀ to the tuning service module 110 . The optimization algorithm in the tuning service module 110 can pass the KPI value y ₃₀ , the transformed KPI value y ₃₀ and the noise variance

Fig. 11 is an exemplary flow chart of the internal processing of the noise diagnosis module when m=60 provided by the embodiment of the present application. When m=60, the tuning object 120 can operate according to the candidate configuration parameter 60 . The performance monitoring module 130 monitors the tuning object 120 according to the preset KPI index, obtains the KPI corresponding to the candidate configuration parameter 60, and sends the KPI corresponding to the candidate configuration parameter 60 to the noise diagnosis module. The noise diagnosis module 140 may execute the process shown in FIG. 11 according to the received KPI.

910. Outlier detection.

It should be understood that the historical performance observation record set when m=60 includes the initial performance observation record set and KPIs corresponding to candidate configuration parameters 1 to 59 after removing abnormal noise. Perform Gaussian fitting on the set of historical performance observation records to obtain the expected value of the candidate configuration parameter 60, and then judge whether the expected value falls within the confidence interval according to the confidence interval constructed in step 520 in Figure 5, and if it falls within the confidence interval, then the candidate configuration The KPI corresponding to the parameter 60 is not abnormal noise. If it falls outside the confidence interval, the KPI corresponding to the candidate configuration parameter 60 is abnormal noise. The abnormal noise is screened out in this way.

It should be understood that when m=60, the number of iterations M preset by the user has already been met. If the KPI corresponding to the candidate configuration parameter 60 is judged to be abnormal noise, then the abnormal noise is discarded, and the tuning service module 110 directly compares the KPIs corresponding to all candidate configuration parameters in the tuning record storage module 150, and selects when the KPI is optimal The corresponding set of candidate configuration parameters is output as the selection of optimal parameters. If the KPI corresponding to the candidate configuration parameter 60 is judged not to be abnormal noise, the noise diagnosis module 140 will output the KPI, and the tuning service module 110 will save the KPI to the tuning record storage module 150, and then record all candidate configuration parameters in 60 rounds The corresponding non-abnormal noise KPIs are compared, and the corresponding set of candidate configuration parameters output when the KPI is optimal is selected as the optimal parameter selection.

The parameter optimization method according to the embodiment of the present application has been described above, and the apparatus and equipment according to the embodiment of the present application will be described below with reference to FIG. 12 and FIG. 13 respectively.

The embodiment of the present application also provides a computer storage medium, the computer storage medium stores program instructions, and when the program is executed, it may include the part of the parameter optimization method in the corresponding embodiments as shown in Fig. 2, 5, 9-11 or all steps.

FIG. 12 is a structural example diagram of a computer device 1000 provided in an embodiment of the present application. The computer device 1000 includes an acquisition module 1010 and a processing module 1020 .

Among them, the obtaining module 1010 is used to obtain candidate configuration parameters, historical performance observation records set and noise variance, and executes 220 and 280 in the method of FIG. 2, 540 in the method of FIG. 5, 730 in the method of FIG. 9, and 730 in the method of FIG. The 830.

The processing module 1020 is used to select the candidate configuration parameters corresponding to the optimal KPI according to the candidate configuration parameters, the historical performance observation record set and the noise variance, and execute the method in FIG. 2 , the method in FIG. 5 , the method in FIG. 9 , and the method in FIG. 10 , some or all of the steps in the method in FIG. 11 .

FIG. 13 is a structural example diagram of another computer device 1300 provided in the embodiment of the present application. The computer device 1300 includes a processor 1302 , a communication interface 1303 and a memory 1304 . One example of computer device 1300 is a chip. Another example of computer apparatus 1300 is a computing device.

The methods disclosed in the foregoing embodiments of the present invention may be applied to the processor 1302 or implemented by the processor 1302 . Processor 1302 can be a central processing unit (central processing unit, CPU), and can also be other general-purpose processors, digital signal processors (digital signal processors, DSP), application specific integrated circuits (application specific integrated circuits, ASICs), on-site Programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or any conventional processor or the like. In the implementation process, each step of the above method may be completed by an integrated logic circuit of hardware in the processor 1302 or instructions in the form of software. Various methods, steps and logic block diagrams disclosed in the embodiments of the present invention may be implemented or executed. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like. The steps of the methods disclosed in the embodiments of the present invention may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor.

Memory 1304 can be volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory. Among them, the non-volatile memory can be read-only memory (read-only memory, ROM), programmable read-only memory (programmable ROM, PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically programmable Erases programmable read-only memory (electrically EPROM, EEPROM) or flash memory. Volatile memory can be random access memory (RAM), which acts as external cache memory. By way of illustration and not limitation, many forms of RAM are available such as static random access memory (static RAM, SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous connection dynamic random access memory (synchlink DRAM, SLDRAM ) and direct memory bus random access memory (direct rambus RAM, DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but not be limited to, these and any other suitable types of memory.

The processor 1302, the memory 1304, and the communication interface 1303 may communicate through a bus. Executable codes are stored in the memory 1304, and the processor 1302 reads the executable codes in the memory 1304 to execute a corresponding method. The memory 1304 may also include an operating system and other software modules required for running processes. The operating system can be LINUX ^TM , UNIX ^TM , WINDOWS ^TM and so on.

For example, the executable code in the memory 1304 is used to implement the method shown in FIG. 2 , and the processor 1302 reads the executable code in the memory 1304 to execute the method shown in FIG. 2 .

In some embodiments of the present application, the disclosed methods may be implemented as computer program instructions encoded in a machine-readable format on a computer-readable storage medium or on other non-transitory media or articles of manufacture. Figure 14 schematically illustrates a conceptual partial view of an example computer program product comprising a computer program for executing a computer process on a computing device, arranged in accordance with at least some embodiments presented herein. In one embodiment, the example computer program product 1400 is provided using a signal bearing medium 1401 . The signal bearing medium 1401 may include one or more program instructions 1402, which may provide the functions or part of the functions described above with respect to the method shown in FIG. 2 when executed by one or more processors. Thus, for example, with reference to the embodiment shown in FIG. 2 , one or more features of 410 through 450 may be undertaken by one or more instructions associated with signal bearing medium 1401 .

In some examples, signal bearing medium 1401 may comprise computer readable medium 1403 such as, but not limited to, a hard drive, compact disc (CD), digital video disc (DVD), digital tape, memory, read-only memory (read only memory) -only memory, ROM) or random access memory (random access memory, RAM) and so on. In some implementations, signal bearing media 1401 may comprise computer recordable media 1404 such as, but not limited to, memory, read/write (R/W) CDs, R/W DVDs, and the like. In some implementations, signal bearing media 1401 may include communication media 1405 such as, but not limited to, digital and/or analog communication media (eg, fiber optic cables, waveguides, wired communication links, wireless communication links, etc.). Thus, for example, signal bearing medium 1401 may be conveyed by a wireless form of communication medium 1405 (eg, a wireless communication medium that complies with the IEEE 802.11 standard or other transmission protocol). One or more program instructions 1402 may be, for example, computer-executable instructions or logic-implemented instructions. In some examples, the aforementioned computing device may be configured to, in response to program instructions 1402 communicated to the computing device via one or more of computer-readable media 1403 , computer-recordable media 1404 , and/or communication media 1405 , Various operations, functions, or actions are provided. It should be understood that the arrangements described herein are for example purposes only. Accordingly, those skilled in the art will appreciate that other arrangements and other elements (e.g., machines, interfaces, functions, sequences, and groups of functions, etc.) can be used instead, and some elements may be omitted altogether depending on the desired result. . In addition, many of the described elements are functional entities that may be implemented as discrete or distributed components, or implemented in conjunction with other components in any suitable combination and location.

Those of ordinary skill in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present application.

Those skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.

If the functions described above are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disc and other media that can store program codes. .

The above is only a specific implementation of the application, but the scope of protection of the application is not limited thereto. Anyone familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the application. Should be covered within the protection scope of this application. Therefore, the protection scope of the present application should be determined by the protection scope of the claims.

Claims

A parameter optimization method, characterized in that, comprising:

Obtain the first candidate configuration parameter of the tuning object;

Running the tuning object according to the first candidate configuration parameter to obtain a first key performance indicator KPI, where the first KPI is determined according to at least one observation indicator;

Acquiring a first set of historical performance observation records and the variance of the first set of historical performance observation records, the first set of historical performance observation records includes a plurality of KPIs obtained by running the tuning object according to historical candidate configuration parameters, the historical candidate The configuration parameter is a candidate configuration parameter determined before acquiring the first candidate configuration parameter;

determining a first expected value according to the first set of historical performance observation records;

determining a noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set;

Determine a second candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object.
The method according to claim 1, wherein, before determining the first expected value according to the first historical performance observation record set, the method further comprises:

It is determined that the first KPI is not an abnormal performance observation.
The method according to claim 2, further comprising:

running the tuning object according to the second candidate configuration parameters to obtain a second KPI;

determining the second KPI as the abnormal performance observation;

Determine a third candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object.
The method according to claim 3, wherein the determining that the first KPI is not an abnormal performance observation value comprises:

get the confidence interval;

determining that the first expected value is within the confidence interval;

The determining that the second KPI is the abnormal performance observation value includes:

Determine a second expected value according to a second set of historical performance observation records, wherein the second set of historical performance observation records includes a plurality of KPIs in the first set of historical performance observation records and the first KPI;

It is determined that the second expected value lies outside the confidence interval.
The method according to claim 4, wherein said obtaining a confidence interval comprises:

Configure the fourth candidate configuration parameter of the tuning object, and repeat the operation G times to obtain the initial performance observation record set, where G is a positive integer greater than 1;

The confidence interval is obtained according to the initial performance observation record set.
The method according to claim 5, characterized in that, according to the first KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, determining the second of the tuning object Before the two candidate configuration parameters, the method also includes:

determining whether the initial set of performance observation records conforms to an approximate Gaussian distribution;

If the initial performance observation record set does not conform to the approximate Gaussian distribution, then determine a transformation parameter, the transformation parameter can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution;

The determining the second candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object includes:

If the set of initial performance observation records conforms to the approximate Gaussian distribution, then determine the first KPI of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Two candidate configuration parameters;

If the set of initial performance observation records does not conform to the approximate Gaussian distribution, then use the transformation parameters to convert the first KPI into a third KPI, according to the first KPI, the third KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, and determine the second candidate configuration parameters of the tuning object.
The method according to claim 5, characterized in that, according to the first KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, determining the second of the tuning object Before the three candidate configuration parameters, the method also includes:

determining whether the initial set of performance observation records conforms to an approximate Gaussian distribution;

If the initial performance observation record set does not conform to the approximate Gaussian distribution, then determine a transformation parameter, the transformation parameter can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution;

The determining the third candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object includes:

If the set of initial performance observation records conforms to the approximate Gaussian distribution, then determine the first KPI of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Three candidate configuration parameters;

If the set of initial performance observation records does not conform to the approximate Gaussian distribution, then use the transformation parameters to convert the first KPI into a fourth KPI, according to the first KPI, the fourth KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, and determine a third candidate configuration parameter of the tuning object.
The method according to any one of claims 1 to 7, wherein the determining the noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set includes : Determine the noise variance according to the following formula:

Wherein, i is the serial number of the candidate configuration parameter of the tuning object,
is the variance of the first historical performance observation record set, E(xi ) is the expected value of the first historical performance observation record set, y i is the first KPI,
is the noise variance.
A computer device, characterized in that the device includes an acquisition module and a processing module,

The obtaining module is used to obtain the first candidate configuration parameter of the tuning object;

The processing module is configured to run the tuning object according to the first candidate configuration parameter to obtain a first key performance indicator KPI, and the first KPI is determined according to at least one observation indicator;

The acquiring module is further configured to acquire a first set of historical performance observation records and a variance of the first set of historical performance observation records, the first set of historical performance observation records includes running the tuning object according to historical candidate configuration parameters A plurality of KPIs, the historical candidate configuration parameters are candidate configuration parameters determined before acquiring the first candidate configuration parameters;

The processing module is further configured to determine a first expected value according to the first set of historical performance observation records;

The processing module is further configured to determine a noise variance according to the first KPI, the first expected value, and the variance of the first historical performance observation record set;

The processing module is further configured to determine a second candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object.
The device according to claim 9, wherein the processing module,

It is also used to determine that the first KPI is not an abnormal performance observation value.
The device according to claim 10, wherein the processing module,

It is also used to run the tuning object according to the second candidate configuration parameter to obtain a second KPI;

The processing module is further configured to determine that the second KPI is the abnormal performance observation value;

The processing module is further configured to determine a third candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object.
The device according to claim 11, wherein the processing module is specifically used for:

get the confidence interval;

determining that the first expected value is within the confidence interval;

The determining that the second KPI is the abnormal performance observation value includes:

Determine a second expected value according to a second set of historical performance observation records, wherein the second set of historical performance observation records includes a plurality of KPIs in the first set of historical performance observation records and the first KPI;

It is determined that the second expected value lies outside the confidence interval.
The device according to claim 12, wherein the processing module is specifically used for:

Configure the fourth candidate configuration parameter of the tuning object, and repeat the operation G times to obtain the initial performance observation record set, where G is a positive integer greater than 1;

The confidence interval is obtained according to the initial performance observation record set.
The device according to claim 13, wherein the processing module is specifically used for:

determining whether the initial set of performance observation records conforms to an approximate Gaussian distribution;

If the initial performance observation record set does not conform to the approximate Gaussian distribution, then determine a transformation parameter, the transformation parameter can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution;

The determining the second candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object includes:

If the set of initial performance observation records conforms to the approximate Gaussian distribution, then determine the first KPI of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Two candidate configuration parameters;

If the set of initial performance observation records does not conform to the approximate Gaussian distribution, then use the transformation parameters to convert the first KPI into a third KPI, according to the first KPI, the third KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, and determine the second candidate configuration parameters of the tuning object.
The device according to claim 13, wherein the processing module is specifically used for:

determining whether the initial set of performance observation records conforms to an approximate Gaussian distribution;

If the initial performance observation record set does not conform to the approximate Gaussian distribution, then determine a transformation parameter, the transformation parameter can convert the initial performance observation record set into a performance observation record set conforming to the approximate Gaussian distribution;

The determining the third candidate configuration parameter of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameter of the tuning object includes:

If the set of initial performance observation records conforms to the approximate Gaussian distribution, then determine the first KPI of the tuning object according to the first KPI, the noise variance, and the type and range of the candidate configuration parameters of the tuning object. Three candidate configuration parameters;

If the set of initial performance observation records does not conform to the approximate Gaussian distribution, then use the transformation parameters to convert the first KPI into a fourth KPI, according to the first KPI, the fourth KPI, the noise variance, the type and range of the candidate configuration parameters of the tuning object, and determine a third candidate configuration parameter of the tuning object.
The device according to any one of claims 9 to 15, wherein the processing module is specifically used for:

The noise variance is determined according to the following formula:

Wherein, i is the serial number of the candidate configuration parameter of the tuning object,
is the variance of the first historical performance observation record set, E(xi ) is the expected value of the first historical performance observation record set, y i is the first KPI,
is the noise variance.
A computer device, characterized in that it comprises: a processor, configured to be coupled to a memory, read and execute instructions and/or program codes in the memory, so as to execute any one of claims 1-8 method described in the item.
A system on a chip, characterized in that it includes: a logic circuit, the logic circuit is used to couple with an input/output interface, and transmit data through the input/output interface, so as to perform the operation described in any one of claims 1-8. described method.
A computer-readable medium, characterized in that the computer-readable medium stores program codes, and when the computer program codes run on a computer, the computer executes the method described in any one of claims 1-8. method.