CN106250306A - A kind of performance prediction method being applicable to enterprise-level O&M automatization platform - Google Patents

Abstract

The invention discloses a kind of performance prediction method being applicable to enterprise-level O&M automatization platform, it comprises the following steps: S1: data input: being acquired data and after pretreatment of O&M automatization platform, sends to performance prediction module；S2: Model Selection: according to account of the history and existing situation, selects the foreseeable calculation of Progressive symmetric erythrokeratodermia；S3: performance prediction calculates: carry out performance prediction calculating according to the calculation of the performance prediction selected；Described performance prediction includes CPU/ memory prediction and disk prediction；S4: set up the evaluation criterion of forecast model, contrasts the predictive value of actual value with forecast model, sets up self study process: when the predictive value of forecast model is unsatisfactory for specification error, according to the prediction model parameters of actual value amendment forecast model.O&M automatization platform is set up load estimation mechanism and algorithm predicts model by the present invention, completes the prediction for resource service conditions such as CPU, internal memory, disks；And prediction algorithm can carry out unrestricted choice according to practical situation.

Description

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform

Technical field

The present invention relates to a kind of performance prediction method being applicable to enterprise-level O&M automatization platform.

Background technology

Along with the propelling of Digitalization in China process, the level of IT application and the degree of the large organization mechanism such as government, enterprise are big Big raising, when the river rises the boat goes up equally for the scale of application system and complexity.Along with the development of business, the scale day of enterprise data center Benefit is huge, occurs in that many new problems, such as in daily O&M:

The most multi-vendor different types of equipment varies, and due to history or technology, tends not to unified acquisition The information during operation of equipment；

B. operate by flow process by hand during fault and investigate, adding the time of troubleshooting；

C. Daily Round Check relies on manual execution, and efficiency is low and easily makes mistakes；

D. a large amount of IT element such as flow process, data, case, warning, event is dispersed in everywhere, forms multiple information island, lacks Few unified and effective management.

The management mode of tradition O&M needs management personnel's manual monitoring system status, daily to occur in application system Management operation carries out manual process, and cost is high, and efficiency is low and lacks real-time, has not been suitable for large-scale application system.Especially Being the Enterprise Application Management scene for height clustered, automatization's operation management mode is essential.

The automatization of so-called operation management refers to by repetitive operation substantial amounts of in daily O&M is (little to simple day Often inspection, configuration change and software are installed and update, the big organizational scheduling to whole changing process) turned by the craft execution in past For automation mechanized operation based on prefabricated management strategy, thus reduce or even eliminate the delay in O&M, it is achieved the fortune of " zero propagation " Dimension.

Based on this, the CPU/ internal memory of O&M automatization platform and the situation of disk are predicted, for follow-up behaviour by existing offer The method making to provide basis, promotes the efficiency of system O&M, reduces cost of labor.

Summary of the invention

It is an object of the invention to overcome the deficiencies in the prior art, it is provided that one is applicable to enterprise-level O&M automatization platform Performance prediction method, the CPU/ internal memory of O&M automatization platform and the situation of disk are predicted, for subsequent operation provide Basis.

It is an object of the invention to be achieved through the following technical solutions: one is applicable to enterprise-level O&M automatization platform Performance prediction method, it comprises the following steps:

S1: data input: being acquired data and after pretreatment of O&M automatization platform, send to performance prediction Module；

S2: Model Selection: according to account of the history and existing situation, selects the foreseeable calculation of Progressive symmetric erythrokeratodermia；

S3: performance prediction calculates: carry out performance prediction calculating according to the calculation of the performance prediction selected；Described property Prediction can include CPU/ memory prediction and disk prediction；

S4: model evaluation: set up the evaluation criterion of forecast model, it is right actual value and the predictive value of forecast model to be carried out Ratio, sets up self study process: when the predictive value of forecast model is unsatisfactory for specification error, according to actual value amendment forecast model Prediction model parameters, it is ensured that select suitable model in specific environment and special time；

The calculation of described CPU/ memory prediction includes prediction based on autoregression model, Similarity matching prediction, base Prediction in weighted least-squares AR model；

The calculation of described disk prediction includes prediction based on unitary linear fit, based on unitary nonlinear fitting Prediction, prediction based on quadratic fit, prediction based on linear weighted function algorithm, prediction based on cubic fit and based on improvement The prediction of MA model；

Described prediction based on autoregression model includes following sub-step:

S3111: according to the CPU/ internal memory load historical data periodically gathered according to time series of step S1 input, build Stand one about the relational model between the load of CPU/ internal memory and time variable:

y_t=φ₀+φ₁y_t-1+φ₂y_t-2+…+φ_py_t-p+e_t；

In formula, y₁～y_tBeing time series, P rank autoregression model shows y in sequence_tBe front p sequence linear combination and The function of error term, φ₀It is constant term, φ₁～φ_pIt is model parameter, e_tBe possess average be 0, variance be the white noise of σ；

S3112: use the relational model set up in step S3111 to calculate what the CPU/ internal memory corresponding to special time loaded Value, and the predictive value that this value is loaded as CPU/ internal memory；

Described Similarity matching is predicted and is included following sub-step:

S3121: build match pattern, the historical data of calculating CPU/ internal memory load interconversion rate on each time point CR；

S3122: search all qualified parallel patterns, by described parallel pattern according to the time apart from present mode Distance is ranked up, and gives the weight that these parallel patterns are different:

$P_{CPU}\left(t_0\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*P_{CPU}\left(t_i\right);$

P in formula_CPU(t_i) represent away from present mode i-th CPU/ internal memory load model, i=0,1 ..., the biggest expression of n, i from The time of present mode is the most remote, wherein when i=0 when, and P_CPU(t₀) represent current CPU/ internal memory load model；α_iRepresent P_CPU (t_iWeight corresponding to), the biggest α of i_iThe least；Simultaneously by stop time point t of the parallel pattern of different weights_iAnd the power of correspondence Weight α_iSave as and gather P:

P={ (t_i,α_i) | i=1,2 ..., n}；

In formula, n represents set sizes, the sum of the parallel pattern i.e. detected；

S3123: calculate predictive value, the parallel pattern during traversal gathers P successively:

(1) t is worked as_i+n_predict≤t₀, i.e. interval between i-th parallel pattern and current time is more than prediction length n_predictTime, to the n after i-th parallel pattern_predictCPU/ internal memory load value on individual time point performs below equation Calculate:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., n_predict；Represent parallel pattern P_iStop time point t_iCPU/ internal memory load value corresponding to kth point afterwards；

(2) t is worked as_i+n_predict＞ t₀, i.e. interval between i-th parallel pattern and current time is less than prediction length n_predictTime, by t_iTo t₀Between CPU/ internal memory load value be multiplied by correspondence weights be added to following t₀-t_iPredictive value, public Formula is as follows:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., t₀-t_i；

Described prediction based on weighted least-squares AR model includes following sub-step:

S3131: take out the CPU/ memory negative of different time points in each time period by SNPM according to the input of step S1 Carry, after obtaining different data, data are carried out difference processing and standardization；

S3132: obtain AIC value corresponding to each predictive value p by AIC criterion, therefrom selects the p corresponding for AIC of minimum It is worth the exponent number as model；

S3133: use LS method to solve the parameter phi of AR model, then choose a weight matrix and parameter is carried out excellent Change, AR model parameter β after being optimized；φ method of least square is obtained, and β method of weighting is improved and obtained；

S3134: bring parameter into expression formula, obtains prediction data:

X_t=φ₁X_n-1+φ₂X_n-2+...+φ_pX_n-p+α_n；

In formula, X is load sequence, α_nFor white noise；

Described prediction based on unitary linear fit includes following sub-step:

S3211: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in linear function by time pretreatment, and obtain utilization rate for sky Data pair, the rate of being set using is the meansigma methods of all utilization rates；

S3212: the some discrete data pair obtained according to step S3211, calculate the undetermined coefficient of linear function；

S3213: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3212；

S3214: the time point needing prediction is carried out the pretreatment identical with step S3211, and is brought into step S3213 The fitting function of structure, i.e. can get the value of utilization rate corresponding to time point；

Described prediction based on unitary nonlinear fitting includes following sub-step:

S3221: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in nonlinear function by time pretreatment, and obtain utilization rate and be Empty data pair, the rate of being set using is the meansigma methods of all utilization rates；Described nonlinear function includes hyperbolic function, index Function and power function；

S3222: the some discrete data pair obtained according to step S3221, calculate the undetermined of each nonlinear function Coefficient；

S3223: bring back in nonlinear function by the undetermined coefficient obtained in step S3222, obtains each matching undetermined Function；

S3224: calculate the degree of fitting of all fitting functions undetermined, selects the function that degree of fitting is little as prediction matching letter Number；

S3225: the time point needing prediction is carried out the pretreatment identical with step S3221, and is brought into step S3224 The prediction fitting function selected, i.e. can get the value of utilization rate corresponding to time point；

Described obtain prediction based on linear weighted function algorithm and include following sub-step:

S3231: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in linear function by time pretreatment, and obtain utilization rate for sky Data pair, the rate of being set using is the meansigma methods of all utilization rates；

S3232: the some discrete data pair obtained according to step S3231, calculate the undetermined coefficient of linear function；

S3233: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3232；

S3234: calculate the utilization rate of historical data according to the fitting function constructed, according to the historical data calculated Utilization rate and the actual value of historical data utilization rate be calculated weights, recalculate the undetermined coefficient of function according to weights, And construct new fitting function；

S3235: the time point needing prediction is carried out the pretreatment in same step S3231, is brought into structure in step S3234 The fitting function made i.e. can obtain the value of utilization rate corresponding to this time；

Described include following sub-step based on the prediction improving MA model:

S3241: the disk utilization rate data obtained according to step S1, to the auto-correlation of disk utilization rate data with partially from phase Close coefficient model is determined rank；

S3242: time series carried out such as stationary test, eliminate singular value, data smoothing process；

S3243: be calculated minimum lag value by formula；

S3244: use the transferring weights method improved to come to each time point and give weights；

S3245: use many new breath RLSs to carry out the parameter estimation of model；

S3246: disk utilization rate is predicted according to the parameter model obtained.

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform also includes an alarm and aid decision Step S5, including following sub-step: monitor the status information of monitored device, believes including current state information and predicted state Breath, when a certain status data value of equipment exceedes set threshold value, alarm module will be entered by the alarm mode set in advance Row alarm, and build knowledge base according to common O&M fault, while making warning information and provide aid decision.

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform also includes data display step S6, The prediction data that pretreated data and step S3 step S1 obtained obtains is shown.

During described step S2 selects, by the way of algorithm pattern/table, show the pluses and minuses of each algorithm, User selects according to practical situation.

Aid decision described in step S2 includes automatic dilatation, including following sub-step:

S50, initial configuration: preserving three configuration files under configuration file catalogue, wherein first file is used for protecting Depositing the essential information of storage device, described storage device essential information includes the IP message file name of storage device and uses institute State the host information filename of storage device, wherein the IP message file name of storage device and the host information of use storage device The filename of the most corresponding second file of the concrete value of filename and the filename of the 3rd file；

The content of second described file includes IP address and the group of port numbers of storage device；

The 3rd described file is for mapping the fdisk or file partition of using storage device in main frame For the device id needed for webservice, whether particular content includes using the IP address of main process equipment of storage device, main frame Use the equipment needed for the file system of storage device or fdisk value and the super equipment of file system correspondence size ID；

S51: read configuration file, obtains device id number；

S52: the disc information that inquiry main process equipment is corresponding；

S53: check disk utilization, and compare the disk partition that configuration file is corresponding with data base: when disk divides District or disk utilization rate have exceeded threshold value, then utilize Webservice interface to carry out dilatation；The most directly terminate.

The invention has the beneficial effects as follows:

(1) present invention O&M automatization platform is set up load estimation mechanism with algorithm predicts model, complete for CPU, The prediction of the resource service condition such as internal memory, disk；And prediction algorithm can carry out unrestricted choice according to practical situation.

(2) combine load estimation result, analyze warning information and provide relevant aid decision, use the side such as script, api interface Formula realizes resource capacity expansion, troubleshooting.By with the ITSM system integration, set up related question knowledge base, to operation maintenance personnel provide Solution.

(3) present invention additionally comprises an algorithm evaluation module, set up prediction algorithm evaluation criterion, by actual value and pre-measuring and calculating The predictive value of method contrasts, and sets up self study process, constantly revises respective model parameter, improves forecasting accuracy, it is ensured that Specific environment and special time select suitable algorithm model.

Accompanying drawing explanation

Fig. 1 is the inventive method flow chart.

Detailed description of the invention

Technical scheme is described in further detail below in conjunction with the accompanying drawings: as it is shown in figure 1, one is applicable to enterprise-level The performance prediction method of O&M automatization platform, it comprises the following steps:

S1: data input: being acquired data and after pretreatment of O&M automatization platform, send to performance prediction Module；

S2: Model Selection: according to account of the history and existing situation, selects the foreseeable calculation of Progressive symmetric erythrokeratodermia；

Specifically, select according to the feature of different models:

For CPU/ memory prediction: such as want to select working stability and low may be selected by based on certainly of time complexity The prediction of regression model；If the aspect ratio of match pattern more apparent (data waveform is bigger), it is possible to select Similarity matching pre- Survey, more preferable data can be obtained；If it is desired to select fitting effect is good, mean absolute error is little, it is possible to select based on adding The prediction of power least square AR model.

Disk is predicted: once, secondary, cubic fit, according to practical situation, three times the most accurately but data calculate Measure bigger.Also having linear weighted function algorithm predicts (carrying out the anti-matching again postponed of actual value), data performance is more preferable.And based on changing Entering the prediction of MA model, reduce lagged value, it is therefore prevented that the impact that data acquisition errors is caused, fitting effect is more preferable, it was predicted that Error is less.

S3: performance prediction calculates: carry out performance prediction calculating according to the calculation of the performance prediction selected；Described property Prediction can include CPU/ memory prediction and disk prediction；

The calculation of described CPU/ memory prediction includes prediction based on autoregression model, Similarity matching prediction, base Prediction in weighted least-squares AR model；

The calculation of described disk prediction includes prediction based on unitary linear fit, based on unitary nonlinear fitting Prediction, prediction based on linear weighted function algorithm, prediction based on quadratic fit, prediction based on cubic fit and based on improvement The prediction of MA model；

Described prediction based on autoregression model includes following sub-step:

S3111: according to the CPU/ internal memory load historical data periodically gathered according to time series of step S1 input, build Stand one about the relational model between the load of CPU/ internal memory and time variable:

y_t=φ₀+φ₁y_t-1+φ₂y_t-2+…+φ_py_t-p+e_t；

In formula, y₁～y_tBeing time series, P rank autoregression model shows y in sequence_tBe front p sequence linear combination and The function of error term, φ₀It is constant term, φ₁～φ_pIt is model parameter, e_tBe possess average be 0, variance be the white noise of σ；

S3112: use the relational model set up in step S3111 to calculate what the CPU/ internal memory corresponding to special time loaded Value, and the predictive value that this value is loaded as CPU/ internal memory；

Autoregression model working stability and there is relatively low time complexity.

Described Similarity matching is predicted and is included following sub-step:

S3121: build match pattern, the historical data of calculating CPU/ internal memory load interconversion rate on each time point CR；

S3122: search all qualified parallel patterns, by described parallel pattern according to the time apart from present mode Distance is ranked up, and gives the weight that these parallel patterns are different:

$P_{CPU}\left(t_0\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*P_{CPU}\left(t_i\right);$

P in formula_CPU(t_i) represent away from present mode i-th CPU/ internal memory load model, i=0,1 ..., the biggest expression of n, i from The time of present mode is the most remote, wherein when i=0 when, and P_CPU(t₀) represent current CPU/ internal memory load model；α_iRepresent P_CPU (t_iWeight corresponding to), the biggest α of i_iThe least；Simultaneously by stop time point t of the parallel pattern of different weights_iAnd the power of correspondence Weight α_iSave as and gather P:

P={ (t_i,α_i) | i=1,2 ..., n}；

In formula, n represents set sizes, the sum of the parallel pattern i.e. detected；

S3123: calculate predictive value, the parallel pattern during traversal gathers P successively:

(1) t is worked as_i+n_predict≤t₀, i.e. interval between i-th parallel pattern and current time is more than prediction length n_predictTime, to the n after i-th parallel pattern_predictCPU/ internal memory load value on individual time point performs below equation Calculate:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., n_predict；Represent parallel pattern P_iStop time point t_iCPU/ internal memory load value corresponding to kth point afterwards；

(2) t is worked as_i+n_predict＞ t₀, i.e. interval between i-th parallel pattern and current time is less than prediction length n_predictTime, by t_iTo t₀Between CPU/ internal memory load value be multiplied by correspondence weights be added to following t₀-t_iPredictive value, public Formula is as follows:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., t₀-t_i。

Specifically, embodiment 1 is the CPU runnability Forecasting Methodology using Similarity matching prediction:

S01 (corresponding step S1): obtain the CPU running state information and CPU self extracted after pretreatment The cpu load historical data of performance information, and calculate the cpu load Load on each timing node_cpu, formula is as follows:

${\mathrm{Load}}_{cpu}=\frac{used\_cpu}{all\_cpu},$

In formula, used_cpu represents that the CPU used, all_cpu represent CPU altogether, and both is with million instructions MIPS per second is unit；

S02 (corresponding step S3121): the cpu load historical data of input is carried out calculating pretreatment, and it is negative to calculate CPU The rate of change CR being loaded on each timing node, formula is as follows:

$CR\left(t_i\right)=\frac{V_{t_i}-V_{t_{i-1}}}{V_{t_{i-1}}}*100\%;$

In formula,For t_iThe cpu load numerical value in moment, i=0,1,2 ..., n；

S03: choose present mode matching length L (i.e. n_predict), structure cpu load historical data and present load mould Formula, the rate of change of cpu load and current rate of change pattern；

S04 (i.e. all qualified parallel patterns of the lookup of step S3122): the cpu load historical data to input Carrying out segmentation, hop count is N, calculates degree of similarity piecemeal, including following sub-step:

S041: the value making variable i is 0；

S042: judge whether the condition that the value of i is less than N meets:

(1) if condition does not meets entrance step S05；

(2) if condition meets, including following sub-step:

A: calculate SD_CR(t_i) and SD_CPU(t_i), wherein SD_CR(t_i) represent in CPU historical record data from t_i-L arrives t_iThis The standard deviation of the cpu load of record, SD in the section time_CPU(t_i) it is the standard deviation of the cpu load of present mode record；

B: judge whether at least one in following two condition is set up, if set up, enters step S043, if not Set up and after then value to i adds 1 operation, return step S042:

Condition 1:SD_CPU(t_i)≤SD_{CPU_THRESHOLD}；

Condition 2:SD_CPU(t_i)≤SD_{CPU_TOLERANCE}And SD_CR(t_i)≤SD_{CR_THRESHOLD}

In formula, SD_{CPU_THRESHOLD}Represent the maximum cpu load standard variance of parallel pattern standard, SD_{CPU_TOLERANCE}Represent The tolerance value of cpu load standard variance, SD_{CR_THRESHOLD}Represent the maximum cpu load rate of change standard side of parallel pattern standard Difference；

S043: by t_iThe SD at place_CR(t_i) and SD_CPU(t_i) add in parallel pattern set Q；

S05 (i.e. step S3122 by described parallel pattern according to distance present mode time distance be ranked up and compose Give weights): the parallel pattern in set Q is ranked up by the distance of distance present mode time, and gives different weights；

The structure set of i.e. step S3122 of S06:(): by stop time point t of the parallel pattern of different weights_iPoint and Respective weights is saved in set P, and formula is as follows:

$P_{CPU}\left(t_0\right)=\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i*P_{CPU}\left(t_i\right),$

In formula, P_CPU(t_i) represent away from present mode i-th cpu load pattern, i=0,1,2 ..., n；The biggest expression of i from The time of present mode is the most remote, α_iRepresent the weight of parallel pattern i；P_CPU(t₀) represent current CPU load pattern；

S07 (i.e. step S3123): traversal set P, calculate each predictive value, compare and obtain final predictive value；

S08 (corresponding step S5 and S6): a series of cpu load predictive values obtained are shown in O&M monitoring management module Show, with time point as abscissa, percentage load describe cpu load-time prediction figure for vertical coordinate, by a period of time in future The loading trends curve chart of CPU represents；Figure will exceed information (time of origin, the load specifying cpu load threshold value simultaneously Value) in the way of alarm, it is sent to system and monitoring personnel, prompting takes measures.

Prediction algorithm based on pattern can change the waveform after predicting according to the waveform in previous scale model Change, the more identical actual waveform measured of waveform therefore predicting out is especially brighter in the feature of current match pattern Time aobvious (i.e. waveform in match pattern have significantly fluctuate), it was predicted that result is more accurate.

Described prediction based on weighted least-squares AR model includes following sub-step:

S3131: take out the CPU/ memory negative of different time points in each time period by SNPM according to the input of step S1 Carry, after obtaining different data, data are carried out difference processing and standardization；

S3132: obtain AIC value corresponding to each predictive value p by AIC criterion, therefrom selects the p corresponding for AIC of minimum It is worth the exponent number as model；

S3133: use LS method to solve the parameter phi of AR model, then choose a weight matrix and parameter is carried out excellent Change, AR model parameter β after being optimized；φ method of least square is obtained, and β method of weighting is improved and obtained；

S3134: bring parameter into expression formula, obtains prediction data:

X_t=φ₁X_n-1+φ₂X_n-2+...+φ_pX_n-p+α_n；

In formula, X is load sequence, α_nFor white noise.

By calculating average absolute value error, comparing traditional AR model, WLS+AR model has good matching to imitate Really, mean absolute error decreases a lot.

The principle of disk prediction is all based on: some discrete function values of certain function known f1, f2 ..., fn}, by adjusting Some undetermined coefficients f in this function whole (λ 1, λ 2 ..., λ m) so that the difference (least square meaning) of this function and known point set Minimum.

Described prediction based on unitary linear fit includes following sub-step:

S3211: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in linear function by time pretreatment, and obtain utilization rate for sky Data pair, the rate of being set using is the meansigma methods of all utilization rates；

S3212: the some discrete data pair obtained according to step S3211, calculate the undetermined coefficient of linear function；

S3213: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3212；

S3214: the time point needing prediction is carried out the pretreatment identical with step S3211, and is brought into step S3213 The fitting function of structure, i.e. can get the value of utilization rate corresponding to time point.

Described prediction based on unitary nonlinear fitting includes following sub-step:

S3221: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in nonlinear function by time pretreatment, and obtain utilization rate and be Empty data pair, the rate of being set using is the meansigma methods of all utilization rates；Described nonlinear function includes hyperbolic function, index Function and power function；

S3222: the some discrete data pair obtained according to step S3221, calculate the undetermined of each nonlinear function Coefficient；

S3223: bring back in nonlinear function by the undetermined coefficient obtained in step S3222, obtains each matching undetermined Function；

S3224: calculate the degree of fitting of all fitting functions undetermined, selects the function that degree of fitting is little as prediction matching letter Number；

S3225: the time point needing prediction is carried out the pretreatment identical with step S3221, and is brought into step S3224 The prediction fitting function selected, i.e. can get the value of utilization rate corresponding to time point；

Described obtain prediction based on linear weighted function algorithm and include following sub-step:

S3231: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, right Data carry out pretreatment, are the data mode that can be updated in linear function by time pretreatment, and obtain utilization rate for sky Data pair, the rate of being set using is the meansigma methods of all utilization rates；

S3232: the some discrete data pair obtained according to step S3231, calculate the undetermined coefficient of linear function；

S3233: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3232；

S3234: calculate the utilization rate of historical data according to the fitting function constructed, according to the historical data calculated Utilization rate and the actual value of historical data utilization rate be calculated weights, recalculate the undetermined coefficient of function according to weights, And construct new fitting function；

S3235: the time point needing prediction is carried out the pretreatment in same step S3231, is brought into structure in step S3234 The fitting function made i.e. can obtain the value of utilization rate corresponding to this time；

Described include following sub-step based on the prediction improving MA model:

S3241: the disk utilization rate data obtained according to step S1, to the auto-correlation of disk utilization rate data with partially from phase Close coefficient model is determined rank；

S3242: time series carried out such as stationary test, eliminate singular value, data smoothing process；

S3243: be calculated minimum lag value by formula；

S3244: use the transferring weights method improved to come to each time point and give weights；

S3245: use many new breath RLSs to carry out the parameter estimation of model；

S3246: disk utilization rate is predicted according to the parameter model obtained.

Traditional WMA model improves, to ensure that it can obtain preferably when processing disk utilization rate data Prediction effect.On the premise of ensureing precision, reduce lagged value, and add transferring weights on tradition WMA model, be used for preventing The impact that prediction is caused by accident error when only gathering data, and combine Pauta criterion, will appear from abnormal time point Weighted value is transferred on other normal time points rather than as missing value is only processed by traditional transferring weights method, Obtain parameter estimation by many new breath RLSs afterwards, obtain more preferable fitting effect.Experiments verify that, permissible Effectively reduce error, the best to the repair efficiency of disk utilization rate sequence hysteresis.

Prediction based on quadratic fit, the step phase predicted with prediction based on unitary linear fit based on cubic fit With, simply fitting function is revised as quadratic function or cubic function.

S4: model evaluation: set up the evaluation criterion of forecast model, it is right actual value and the predictive value of forecast model to be carried out Ratio, sets up self study process: when the predictive value of forecast model is unsatisfactory for specification error, according to actual value amendment forecast model Prediction model parameters, it is ensured that select suitable model in specific environment and special time；

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform also includes an alarm and aid decision Step S5, including following sub-step: monitor the status information of monitored device, believes including current state information and predicted state Breath, when a certain status data value of equipment exceedes set threshold value, alarm module will be entered by the alarm mode set in advance Row alarm, and build knowledge base according to common O&M fault, while making warning information and provide aid decision.

Described aid decision includes:

(1) for each most common failure and warning, system has default treatment mechanism, such as:

Finding that disk to be expired, system is automatically obtained deletion garbage files and dilatation function；

When CPU, memory usage are too high, kill engineering noise process；

When a certain server node residual memory space is not enough, and when other node storage space are sufficient, it is achieved resource Migrate, reach to store load balance.

(2) for each most common failure and warning, system provides a user with multiple treatment mechanism, and user can select to write from memory Recognize treatment mechanism.

(3) for the higher warning of rank and fault, system will not enter default treatment mechanism, will inform the user that, by with Family determines treatment mechanism.

Webservice interface is used to carry out dilatation for automatic dilatation.Premise is the utilization rate of the disk of the equipment of main frame Exceed threshold value.Specifically:

S50, initial configuration: at configuration file catalogue $ ROOT/nariweb/src/main/resources/ Under macrosan, preserve three configuration files, be respectively DeviceMappingLunInfo.xml, Macrosan.properties and macrosan.xml.

Wherein macrosan.properties is for preserving the essential information of storage device, and described storage device is basic Information includes IP message file name MacrosanDeviceInfo of storage device and uses the host information of described storage device Filename HostUseMacrosanInfo, wherein the IP message file name of storage device and the host information of use storage device The filename macrosan.xml of the most corresponding second file of the concrete value of filename and the filename of the 3rd file DeviceMappingLunInfo.xml；That is:

MacrosanDevicesInfo=macrosan.xml；

HostUseMacroSanInfo=DeviceMappingLunInfo.xml.

The content of second described file macrosan.xml includes IP address and the group of port numbers of storage device；

File is the form of XML file, when having new storage device to add when, adds storage in this document and sets Standby address.Please note the content of interpolation in addition to IP address, http to be added: // head and port numbers, concrete form is Http:// IP address: port numbers such as, will add the storage device that ip address is " 10.144.19.172 ", and we will Under this document, add content.I.e.

<ip>http://10.144.19.172:9090</ip>。

The 3rd described file DeviceMappingLunInfo.xml is for using the hard of storage device in main frame Dish subregion or file partition are mapped as the device id needed for webservice, and particular content includes: 1,<hostip>label In, use the IP address of the main process equipment of storage device；2, in<hostsection>label, whether main frame uses storage The file system of equipment or fdisk value；3, needed for the file system super equipment of correspondence size in<deviceid>label Device id；That is:

<hostip>10.144.9.76</hostip>

<deviceid>600B342EE2823DAD9790D902ED0000DD</deviceid>

<hostsection>/</hostsection>

S51: read configuration file, obtains device id number；

S52: the disc information that inquiry main process equipment is corresponding；

S53: check disk utilization, and compare the disk partition that configuration file is corresponding with data base: when disk divides District or disk utilization rate have exceeded threshold value, then utilize Webservice interface to carry out dilatation；The most directly terminate.

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform also includes data display step S6, The prediction data that pretreated data and step S3 step S1 obtained obtains is shown.

During described step S2 selects, by the way of algorithm pattern/table, show the pluses and minuses of each algorithm, User selects according to practical situation.

Claims (3)

1. the performance prediction method being applicable to enterprise-level O&M automatization platform, it is characterised in that: it comprises the following steps:

S1: data input: being acquired data and after pretreatment of O&M automatization platform, send to performance prediction mould Block；

S2: Model Selection: according to account of the history and existing situation, selects the foreseeable calculation of Progressive symmetric erythrokeratodermia；

S3: performance prediction calculates: carry out performance prediction calculating according to the calculation of the performance prediction selected；Described performance is pre- Survey includes CPU/ memory prediction and disk prediction；

S4: model evaluation: set up the evaluation criterion of forecast model, contrasts the predictive value of actual value with forecast model, builds Vertical self study process: when the predictive value of forecast model is unsatisfactory for specification error, according to the prediction of actual value amendment forecast model Model parameter, it is ensured that select suitable model in specific environment and special time；

The calculation of described CPU/ memory prediction includes that prediction based on autoregression model, Similarity matching are predicted, based on adding The prediction of power least square AR model；

The calculation of described disk prediction includes prediction based on unitary linear fit, based on unitary nonlinear fitting pre- Survey, prediction based on quadratic fit, prediction based on linear weighted function algorithm, prediction based on cubic fit and based on improve MA mould The prediction of type；

Described prediction based on autoregression model includes following sub-step:

S3111: according to the CPU/ internal memory load historical data periodically gathered according to time series of step S1 input, set up one Individual about the relational model between the load of CPU/ internal memory and time variable:

y_t=φ₀+φ₁y_t-1+φ₂y_t-2+…+φ_py_t-p+e_t；

In formula, y₁～y_tBeing time series, P rank autoregression model shows y in sequence_tIt is linear combination and the error of front p sequence The function of item, φ₀It is constant term, φ₁～φ_pIt is model parameter, e_tBe possess average be 0, variance be the white noise of σ；

S3112: use the relational model set up in step S3111 to calculate the value of the CPU/ internal memory load corresponding to special time, And the predictive value that this value loaded as CPU/ internal memory；

Described Similarity matching is predicted and is included following sub-step:

S3121: build match pattern, the historical data of calculating CPU/ internal memory load interconversion rate CR on each time point；

S3122: search all qualified parallel patterns, described parallel pattern is far and near according to the time of distance present mode It is ranked up, and gives the weight that these parallel patterns are different:

$P_{CPU}\left(t_0\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i*P_{CPU}\left(t_i\right);$

P in formula_CPU(t_i) represent away from present mode i-th CPU/ internal memory load model, i=0,1 ..., the biggest expression of n, i is from currently The time of pattern is the most remote, wherein when i=0 when, and P_CPU(t₀) represent current CPU/ internal memory load model；α_iRepresent P_CPU(t_i) Corresponding weight, the biggest α of i_iThe least；Simultaneously by stop time point t of the parallel pattern of different weights_iAnd the weight α of correspondence_i Save as and gather P:

P={ (t_i,α_i) | i=1,2 ..., n}；

In formula, n represents set sizes, the sum of the parallel pattern i.e. detected；

S3123: calculate predictive value, the parallel pattern during traversal gathers P successively:

(1) t is worked as_i+n_predict≤t₀, i.e. interval between i-th parallel pattern and current time is more than prediction length n_predictTime, To the n after i-th parallel pattern_predictThe calculating of the CPU/ internal memory load value execution below equation on individual time point:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., n_predict；V_Pi(t_i+ k) represent parallel pattern P_i's Stop time point t_iCPU/ internal memory load value corresponding to kth point afterwards；

(2) t is worked as_i+n_predict＞ t₀, i.e. interval between i-th parallel pattern and current time is less than prediction length n_predictTime, By t_iTo t₀Between CPU/ internal memory load value be multiplied by correspondence weights be added to following t₀-t_iPredictive value, formula is as follows:

$V_{prediction}\left(k\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i*V_{P_i}(t_i+k);$

In formula, V_predictionThe predictive value of (k) expression kth step, k=1,2 ..., t₀-t_i；

Described prediction based on weighted least-squares AR model includes following sub-step:

S3131: taking out the CPU/ internal memory of different time points in each time period by SNPM and load according to the input of step S1, After different data, data are carried out difference processing and standardization；

S3132: obtain AIC value corresponding to each predictive value p by AIC criterion, the p value corresponding for AIC therefrom selecting minimum is made Exponent number for model；

S3133: use LS method to solve the parameter phi of AR model, then choose a weight matrix and parameter is optimized, AR model parameter β after being optimized；φ method of least square is obtained, and β method of weighting is improved and obtained；

S3134: bring parameter into expression formula, obtains prediction data:

X_t=φ₁X_n-1+φ₂X_n-2+...+φ_pX_n-p+α_n；

In formula, X is load sequence, α_nFor white noise；

Described prediction based on unitary linear fit includes following sub-step:

S3211: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, to data Carry out pretreatment, be the data mode that can be updated in linear function by time pretreatment, and obtain the number that utilization rate is sky According to right, the rate of being set using is the meansigma methods of all utilization rates；

S3212: the some discrete data pair obtained according to step S3211, calculate the undetermined coefficient of linear function；

S3213: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3212；

S3214: the time point needing prediction is carried out the pretreatment identical with step S3211, and is brought into step S3213 structure Fitting function, i.e. can get the value of utilization rate corresponding to time point；

Described prediction based on unitary nonlinear fitting includes following sub-step:

S3221: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, to data Carry out pretreatment, be the data mode that can be updated in nonlinear function by time pretreatment, and to obtain utilization rate be empty Data pair, the rate of being set using is the meansigma methods of all utilization rates；Described nonlinear function includes hyperbolic function, exponential function And power function；

S3222: the some discrete data pair obtained according to step S3221, calculate the undetermined coefficient of each nonlinear function；

S3223: bring back in nonlinear function by the undetermined coefficient obtained in step S3222, obtains each fitting function undetermined；

S3224: calculate the degree of fitting of all fitting functions undetermined, selects the function that degree of fitting is little as prediction fitting function；

S3225: the time point needing prediction is carried out the pretreatment identical with step S3221, and is brought into the selection of step S3224 Prediction fitting function, i.e. can get the value of utilization rate corresponding to time point；

Described obtain prediction based on linear weighted function algorithm and include following sub-step:

S3231: the time obtained according to step S1 and the time pair of disk utilization rate, formation sequence in chronological order, to data Carry out pretreatment, be the data mode that can be updated in linear function by time pretreatment, and obtain the number that utilization rate is sky According to right, the rate of being set using is the meansigma methods of all utilization rates；

S3232: the some discrete data pair obtained according to step S3231, calculate the undetermined coefficient of linear function；

S3233: bring back to, in linear function, obtain fitting function by the undetermined coefficient obtained in step S3232；

S3234: calculate the utilization rate of historical data according to the fitting function constructed, according to making of the historical data calculated It is calculated weights with the actual value of rate and historical data utilization rate, recalculates the undetermined coefficient of function, and structure according to weights Make new fitting function；

S3235: the time point needing prediction is carried out the pretreatment in same step S3231, is brought in step S3234 structure Fitting function i.e. can obtain the value of utilization rate corresponding to this time；

Described include following sub-step based on the prediction improving MA model:

S3241: the disk utilization rate data obtained according to step S1, auto-correlation and the partial autocorrelation system to disk utilization rate data Model is determined rank by number；

S3242: time series carried out such as stationary test, eliminate singular value, data smoothing process；

S3243: be calculated minimum lag value by formula；

S3244: use the transferring weights method improved to come to each time point and give weights；

S3245: use many new breath RLSs to carry out the parameter estimation of model；

S3246: disk utilization rate is predicted according to the parameter model obtained；

The selection of step S2, user selects, wherein according to the feature of different models:

For CPU/ memory prediction: such as want to select working stability and low may be selected by based on autoregression of time complexity The prediction of model；If the aspect ratio of match pattern more apparent (data waveform is bigger), it is possible to select Similarity matching prediction, can To obtain more preferable data；If it is desired to select fitting effect is good, mean absolute error is little, it is possible to select based on weighting minimum Two predictions taking advantage of AR model；

For disk predict: once, secondary, cubic fit, according to practical situation, once fitting data amount of calculation minimum is still Accuracy is relatively low, the most accurately but data amount of calculation is bigger in cubic fit, and quadratic fit is compromised；Linear weighted function algorithm predicts (is entered The anti-matching again postponed of row actual value), data performance is more preferable；Prediction based on improvement MA model, reduces lagged value, it is therefore prevented that The impact that data acquisition errors is caused, fitting effect is more preferable, it was predicted that error is less；

Described method also includes an alarm and aid decision step S5, including following sub-step: monitor monitored device Status information, including current state information and predicted state information, when a certain status data value of equipment exceedes set threshold value Time, alarm module will be alerted by the alarm mode set in advance, and builds knowledge base according to common O&M fault, While making warning information and provide aid decision；

Aid decision described in step S5 includes automatic dilatation, including following sub-step:

S50, initial configuration: preserving three configuration files under configuration file catalogue, wherein first file is deposited for preservation The essential information of storage equipment, described storage device essential information includes that the IP message file name of storage device is deposited described in using The host information filename of storage equipment, wherein the IP message file name of storage device and the host information file of use storage device The filename of the most corresponding second file of the concrete value of name and the filename of the 3rd file；

The content of second described file includes IP address and the group of port numbers of storage device；

The 3rd described file is for being mapped as the fdisk or file partition of using storage device in main frame Device id needed for webservice, particular content includes using whether the IP address of main process equipment of storage device, main frame make With the device id needed for the file system of storage device or fdisk value and the super equipment of file system correspondence size；

S51: read configuration file, obtains device id number；

S52: the disc information that inquiry main process equipment is corresponding；

S53: check disk utilization, and compare the disk partition that configuration file is corresponding with data base: when disk partition or Person's disk utilization rate has exceeded threshold value, then utilize Webservice interface to carry out dilatation；The most directly terminate.

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform the most according to claim 1, its feature It is: also include data display step S6, the prediction that pretreated data and step S3 step S1 obtained obtains Data are shown.

A kind of performance prediction method being applicable to enterprise-level O&M automatization platform the most according to claim 1, its feature It is: during described step S2 selects, by the way of algorithm pattern/table, shows the pluses and minuses of each algorithm, use Family selects according to practical situation.