CN102739469B - Web service response time predicting method based on time sequence - Google Patents

Abstract

The invention discloses a web service response time predicting method based on increment steady-state analysis, belonging to the field of software performance prediction. The method comprises the following steps: obtaining a sequence of a series of measurement values of network service response time according to a fixed unit time interval; subsequently defining and calculating observation window response time increment types; calculating a transition probability matrix among the types; then calculating a steady-state distribution probability of the response time increment based on the transition probability matrix; and finally adding a calculated increment expected value to a response time measurement value at the t moment so as to obtain the prediction value of the response time at the t+1 moment. With the adoption of the method, the defects that traditional Web prediction mold and method depend too much on SLA (Service-Level-Agreement) specifications and hypothesis parameters are constant and unchanged are avoided; and mold support and analysis means are provided for WEB service credibility research.

Description

A kind of based on seasonal effect in time series Web service response time Forecasting Methodology

Technical field

The invention belongs to the field of software performance prediction, especially relate to a kind of based on seasonal effect in time series Web service response time Forecasting Methodology.

Background technology

Web service plays more and more important role as a kind of novel distributed component object model in the field such as ecommerce, Enterprise information integration, it is considered to one of technology most crucial in service-oriented computing framework (SOA), by Web service combination technology, single Web service combination is got up to become complete application.But Web service operates in the network environment of dynamic change, its response time value is also in frequent variations, the network environment of this dynamic change will affect the run time behaviour of Web service, and then cause the change of whole Web service combination performance, therefore, carrying out to web services the key technology that performance prediction is the application system quality ensured based on service, is also Optimized Service combination important method.In Web service performance, most important two indices is response time and throughput, and the former system completes the time needed for Given task, and the latter represents that system completes the number of times of Given task within the unit interval.

Although academia and industrial quarters propose a series of model for analysis and prediction Web service performance and method, but most of method is mainly with based on SLA(Service-level-agreement, service-level agreement) static analysis be Main Means, service-level agreement is about a contract between Internet service provider and client, there is defined COS, the term such as service quality and customer payment, the Smallest connection bandwidth of the serviced component that it constrains service provider to provide in quantity, maximum packet loss, linkage fault rate equals performance-relevant parameter, based on the Static Performance Analysis of SLA, the limit value of the direct exactly above-mentioned parameter retrained with service-level agreement is for mode input, and suppose that performance parameter constant is constant, the actual performance of analysis and prediction Web service assembly, there is very large defect in the method for this static analysis: due to the network that relies on when Web service runs and system environments fast changing, the various parameter amount affecting performance is as bandwidth, packet loss, linkage fault rate, message field length etc., can not maintain invariable, there is great deviation in the invariable hypothesis of this and performance parameter, because the parameter provided in service-level agreement is upper and lower bound value, value corresponding when Web service runs higher or lower than limit value, thus may cause the performance prediction method based on service-level agreement to be over-evaluated or underestimates actual performance.

In order to make up the deficiency of said method, take into full account the dynamic of Web service performance under true environment, the present invention, in further investigation performance test value time series data, analyzes on the basis of its Long-term change trend, proposes a kind of based on seasonal effect in time series Web service response time Forecasting Methodology.

Summary of the invention

Because the above-mentioned defect of prior art, technical problem to be solved by this invention is to provide one Web service more accurately response time Forecasting Methodology.

For achieving the above object, the invention provides a kind of based on seasonal effect in time series Web service response time Forecasting Methodology, perform according to the following steps:

Step one: the response time data sequence obtaining web services according to fixing unit interval;

Setting response time data sequence has the surveying record of t response time, and described response time data sequence is r _t(i), { r _t(i) | 1≤i≤t, 1≤t < ∞ }; The response time data sequence of described acquisition web services is obtained by the url address test of SOAP UI test platform to web services;

Step 2: the average increment calculating the continuous response time;

Being compiled k continuous print response time is one group, and k is positive integer, and whole response time sequence is divided into individual group, every group is considered as one and investigates window; An xth y response time investigated in window is r _t((x-1) × k+y); A setting xth response time average increment investigated in window is AINC(x);

$\mathrm{AINC}\left(x\right)=\frac{\underset{s=1}{\overset{k-1}{\mathrm{\&Sigma;}}}{r}_t((x-1)\&times;k+s+1)-r_t((x-1)\&times;k+s)}{k},$

Step 3: definition response time average increment class transition probability between calculated response time average increment class;

the value of response time average increment minimum in individual response time average increment is MIN,

the value of response time average increment maximum in individual response time average increment is MAX,

The interval of MIN to MX is divided into p class, p is positive integer; The response time average increment that setting xth is investigated in window is map (x) to the mapping function that l is classified, 1≤l≤p; And if only if $\mathrm{MIN}+\frac{\mathrm{MAX}-\mathrm{MIN}}{p}\&times;(l-1)\&le;\mathrm{AINC}\left(i\right)<\mathrm{MIN}+\frac{\mathrm{MAX}-\mathrm{MIN}}{p}\&times;l$ Time, map (x)=l;

The matrix of transition probabilities in the interval of setting MIN to MAX between m class and the n-th class is TRAN(m, n), m, n are positive integer;

$\mathrm{ICOUNT}(m,l)=\left\{\begin{array}{cc}1& \mathrm{ifmap}\left(l\right)=m\\ 0& \mathrm{else}\end{array}\right.;$

$\mathrm{IJCOUNT}(m,n,l)=\left\{\begin{array}{cc}1& \mathrm{ifmap}(l+1)=\mathrm{nandmap}\left(l\right)=m\\ 0& \mathrm{else}\end{array}\right.;$

Step 4: under computing system balance, response time average increment is in the probability of each response time average increment class;

π is under system balancing, and response time average increment is in the probability vector of h response time average increment class; H the component of π is π (h), 1≤h≤p; π (h)=π (h) × TRAN(m, n);

Step 5: the response time in prediction t+1 moment;

The desired value of the response time average increment in setting poised state situation is EAINC;

$\mathrm{EAINC}=\underset{h=1}{\overset{p}{\mathrm{\&Sigma;}}}\mathrm{\&pi;}\left(h\right)\&times;(\mathrm{MIN}+\frac{(\mathrm{MAX}-\mathrm{MIN})\&times;(h-1)}{p}+\frac{\mathrm{MAX}-\mathrm{MIN}}{2p});$ The response time predicted value in setting t+1 moment is PRT; Calculate $\mathrm{EAINC}=\frac{\underset{w=1}{\overset{k-2}{\mathrm{\&Sigma;}}}{r}_t(t+1-w)-r_t(t-w)+(\mathrm{PRT}-r_t\left(t\right))}{k},$ Obtain the response time value that PRT just can obtain needing prediction accurately.

The invention has the beneficial effects as follows: the present invention is in further investigation performance test value time series data, the basis analyzing its Long-term change trend realizes, avoid in traditional web performance prediction model and method and too rely on SLA specification and suppose the deficiency that performance parameter constant is constant, can support and analysis means for the research of web services credibility supplies a model, predictablity rate and precision can greatly be improved.

Accompanying drawing explanation

Fig. 1 is schematic flow sheet of the present invention.

Fig. 2 is response time measured data schematic diagram.

Fig. 3 be response time predicted value with actual value compare schematic diagram.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described:

As shown in Figure 1, a kind of based on seasonal effect in time series Web service response time Forecasting Methodology, perform according to the following steps:

Step one: the response time data sequence obtaining web services according to fixing unit interval;

Setting response time data sequence has the surveying record of t response time, and described response time data sequence is r _t(i), { r _t(i) | 1≤i≤t, 1≤t < ∞ }, the response time data sequence of described acquisition web services is obtained by the url address test of SOAP UI test platform to web services.

Step 2: the average increment calculating the continuous response time;

Being compiled k continuous print response time is one group, and k is positive integer, and whole response time sequence is divided into individual group, every group is considered as one and investigates window; An xth y response time investigated in window is r _t((x-1) × k+y); A setting xth response time average increment investigated in window is AINC(x); $\mathrm{AINC}\left(x\right)=\frac{\underset{s=1}{\overset{k-1}{\mathrm{\&Sigma;}}}{r}_t((x-1)\&times;k+s+1)-r_t((x-1)\&times;k+s)}{k},$

Step 3: definition response time average increment class transition probability between calculated response time average increment class;

the value of response time average increment minimum in individual response time average increment is MIN,

the value of response time average increment maximum in individual response time average increment is MAX,

The interval of MIN to MAX is divided into p class, p is positive integer; The response time average increment that setting xth is investigated in window is map (x) to the mapping function that l is classified, 1≤l≤p; And if only if $\mathrm{MIN}+\frac{\mathrm{MAX}-\mathrm{MIN}}{p}\&times;(l-1)\&le;\mathrm{AINC}\left(i\right)<\mathrm{MIN}+\frac{\mathrm{MAX}-\mathrm{MIN}}{p}\&times;l$ Time, map (x)=l;

The matrix of transition probabilities in the interval of setting MIN to MAX between m class and the n-th class is TRAN(m, n), m, n are positive integer;

$\mathrm{ICOUNT}(m,l)=\left\{\begin{array}{cc}1& \mathrm{ifmap}\left(l\right)=m\\ 0& \mathrm{else}\end{array}\right.;$

$\mathrm{IJCOUNT}(m,n,l)=\left\{\begin{array}{cc}1& \mathrm{ifmap}(l+1)=\mathrm{nandmap}\left(l\right)=m\\ 0& \mathrm{else}\end{array}\right..$

Step 4: under computing system balance, response time average increment is in the probability of each response time average increment class;

π is under system balancing, and response time average increment is in the probability vector of h response time average increment class; H the component of π is π (h), 1≤h≤p; π (h)=π (h) × TRAN(m, n), concrete

$\mathrm{\&pi;}\left(1\right)=\mathrm{\&pi;}\left(1\right)\&times;\left[\begin{array}{cccc}\mathrm{TRAN}\left(\mathrm{1,1}\right)& \mathrm{TRAN}\left(\mathrm{1,2}\right)& ....& \mathrm{TRAN}(p,1)\\ \mathrm{TRAN}(2,1)& \mathrm{TRAN}\left(\mathrm{2,2}\right)& ....& \mathrm{TRAN}(p,2)\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \mathrm{TRAN}(p-1,p-1)& \&CenterDot;\\ \mathrm{TRAN}(p,1)& \mathrm{TRAN}(p,2)& ....& \mathrm{TRAN}(p,p)\end{array}\right],$

$\mathrm{\&pi;}\left(2\right)=\mathrm{\&pi;}\left(2\right)\&times;\left[\begin{array}{cccc}\mathrm{TRAN}\left(\mathrm{1,1}\right)& \mathrm{TRAN}\left(\mathrm{1,2}\right)& ....& \mathrm{TRAN}(p,1)\\ \mathrm{TRAN}(2,1)& \mathrm{TRAN}\left(\mathrm{2,2}\right)& ....& \mathrm{TRAN}(p,2)\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \mathrm{TRAN}(p-1,p-1)& \&CenterDot;\\ \mathrm{TRAN}(p,1)& \mathrm{TRAN}(p,2)& ....& \mathrm{TRAN}(p,p)\end{array}\right],......,$

$\mathrm{\&pi;}\left(p\right)=\mathrm{\&pi;}\left(p\right)\&times;\left[\begin{array}{cccc}\mathrm{TRAN}\left(\mathrm{1,1}\right)& \mathrm{TRAN}\left(\mathrm{1,2}\right)& ....& \mathrm{TRAN}(p,1)\\ \mathrm{TRAN}(2,1)& \mathrm{TRAN}\left(\mathrm{2,2}\right)& ....& \mathrm{TRAN}(p,2)\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& & \&CenterDot;\\ \&CenterDot;& \&CenterDot;& \mathrm{TRAN}(p-1,p-1)& \&CenterDot;\\ \mathrm{TRAN}(p,1)& \mathrm{TRAN}(p,2)& ....& \mathrm{TRAN}(p,p)\end{array}\right].$

Step 5: the response time in prediction t+1 moment;

The desired value of the response time average increment in setting poised state situation is EAINC;

$\mathrm{EAINC}=\underset{h=1}{\overset{p}{\mathrm{\&Sigma;}}}\mathrm{\&pi;}\left(h\right)\&times;(\mathrm{MIN}+\frac{(\mathrm{MAX}-\mathrm{MIN})\&times;(h-1)}{p}+\frac{\mathrm{MAX}-\mathrm{MIN}}{2p});$ The response time predicted value in setting t+1 moment is PRT; $\mathrm{EAINC}=\frac{\underset{w=1}{\overset{k-2}{\mathrm{\&Sigma;}}}{r}_t(t+1-w)-r_t(t-w)+(\mathrm{PRT}-r_t\left(t\right))}{k},$ Can by calculating

$\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}\mathrm{\&pi;}\left(i\right)\&times;(\mathrm{MIN}+\frac{(\mathrm{MAX}-\mathrm{MIN})\&times;(i-1)}{p}+\frac{\mathrm{MAX}-\mathrm{MIN}}{2p})=\frac{\underset{w=1}{\overset{k-2}{\mathrm{\&Sigma;}}}{r}_t(t+1-w)-r_t(t-w)+(\mathrm{PRT}-r_t\left(t\right))}{k}$

Draw and obtain PRT, extrapolate the response time predicted value in t+1 moment thus.

As Fig. 1, shown in Fig. 2, in order to verify the correctness of method and accuracy, be configured to the INTEL i5-760 processor of 2.8G, the PC inside saving as 4G runs and adopts SOAPUI test platform to provide the example of the Web service of Weather information to test to one, the URL address of described web services example is http://www.webservicex.net/globalweather.asmx WSDL, the time started of test is 10: 30: 0 morning of on October 17th, 2011, test interval is 250 milliseconds, continuous acquisition 128 response time test values, wherein front 100 data are used for model foundation, remaining 28 response time test values recorded from 10: 30: 25 morning of on October 17th, 2011 are for modelling verification.

Optimum configurations in model is: k=5, classification quantity p=8, according to the method that the present invention provides, 10: 30: 25 morning of on October 17th, 2011, later response time predicted value was shown in figure 3 with the result of other three model predication values, in Fig. 3, lines 1 are the time series forecasting value adopting method of the present invention to obtain, lines 2 are the time series forecasting value adopting ARMA time series models to obtain, lines 3 are the time series forecasting value adopting the ARMA-O model eliminating outlier to obtain, lines 4 are the time series forecasting value adopting two quantitative prediction model DQ to obtain, lines 5 are actual value, the time series forecasting value that the inventive method obtains and actual value average error rate are 24.4%, the time series forecasting value adopting ARMA time series models to obtain and actual value average error rate are 30.1%, the time series forecasting value adopting the ARMA-O model eliminating outlier to obtain and actual value mean error are 33.6%, the time series forecasting value adopting two quantitative prediction model DQ to obtain and the error of actual value are 42.8%, visible, adopt the method that the present invention proposes, Web service performance prediction achieves better precision.

More than describe preferred embodiment of the present invention in detail.Should be appreciated that those of ordinary skill in the art just design according to the present invention can make many modifications and variations without the need to creative work.Therefore, all technical staff in the art, all should by the determined protection range of claims under this invention's idea on the basis of existing technology by the available technical scheme of logical analysis, reasoning, or a limited experiment.

Claims (2)

1., based on a seasonal effect in time series Web service response time Forecasting Methodology, it is characterized in that performing according to the following steps:

Step one: the response time data sequence obtaining web services according to fixing unit interval;

Setting response time data sequence has the surveying record of t response time, and described response time data sequence is r _t(i); { r _t(i) | 1≤i≤t, 1≤t < ∞ };

Step 2: the average increment calculating the continuous response time;

Being compiled k continuous print response time is one group, and k is positive integer, and whole response time sequence is divided into individual group, every group is considered as one and investigates window; An xth y response time investigated in window is r _t((x-1) × k+y); A setting xth response time average increment investigated in window is AINC (x);

Step 3: definition response time average increment class transition probability between calculated response time average increment class;

the value of response time average increment minimum in individual response time average increment is MIN,

the value of response time average increment maximum in individual response time average increment is MAX,

The interval of MIN to MAX is divided into p class, p is positive integer; Response time average increment in setting xth investigation window is to the lthe mapping function of individual classification is map (x), 1≤ l≤ p; And if only if time, map (x)= l;

The matrix of transition probabilities in the interval of setting MIN to MAX between m class and the n-th class is TRAN (m, n), m, n is positive integer;

Step 4: under computing system balance, response time average increment is in the probability of each response time average increment class;

π is under system balancing, and response time average increment is in the probability vector of h response time average increment class; H the component of π is π (h), 1≤h≤p; π (h)=π (h) × TRAN (m, n);

Step 5: the response time in prediction t+1 moment;

The desired value of the response time average increment in setting poised state situation is EAINC;

the response time predicted value in setting t+1 moment is PRT; calculate obtain PRT.

2. as claimed in claim 1 a kind of based on seasonal effect in time series Web service response time Forecasting Methodology, it is characterized in that: the response time data sequence of described acquisition web services is obtained by the url address test of SOAP UI test platform to web services.