CN109325266B - Response time distribution prediction method for online cloud service - Google Patents

Abstract

The invention relates to the technical field of cloud computing, and aims to solve the problem of response time distribution prediction of real loads of online cloud services, thereby providing guidance for deployment of cloud platform infrastructure and applications. Therefore, the technical scheme adopted by the invention is that the response time distribution prediction method for the online cloud service is combined with the access load prediction method and the tail delay analysis method to realize the prediction of the response time distribution law of the online cloud service, wherein the access load prediction result is used for analyzing the input parameters of the tail delay of the seat. The method is mainly applied to cloud computing occasions.

Description

Response time distribution prediction method for online cloud services

Technical field

The present invention relates to the field of cloud computing technology, in particular to the field of software-defined cloud platform service performance prediction. Specifically, it relates to a response time distribution prediction method for online cloud services.

Background technique

In a shared cloud computing environment, competition for physical resources (such as CPU, Cache, memory bandwidth and other resources) between multiple applications running on the same physical machine leads to performance degradation, and the delay may reach the level of a single application running exclusively on the physical machine. 600 times the delay. In order to overcome performance interference caused by resource competition and ensure service quality, service providers have invested huge hardware costs and operation and maintenance manpower costs.

"A service with poor experience is a non-existent service" - the long response time of online cloud services is close to the impact of service unavailability on users. Therefore, how to reduce the tail delay of online cloud services and improve user experience is an important challenge faced by cloud computing services. Based on the analysis of the operating characteristics of each cloud service, discovering its performance bottleneck is the key to improving the user experience of cloud services. However, due to the complexity of the cloud infrastructure architecture and the huge scale of application deployment, effective online cloud service analysis models and Methods are still very scarce. In addition, most existing works focus on evaluating average performance metrics, such as average response time, average resource utilization, etc., while few studies consider modeling and analyzing the delay distribution of online cloud services. Therefore, it is necessary to design an effective online cloud service analysis model.

Contents of the invention

In order to overcome the shortcomings of the existing technology, the present invention aims to solve the problem of response time distribution prediction of real load of online cloud services, thereby providing guidance for the deployment of cloud platform infrastructure and applications. To this end, the technical solution adopted by the present invention is to predict the response time distribution of online cloud services by combining the access load prediction method and the tail delay analysis method to predict the response time distribution of online cloud services. The access load prediction method Input parameters for result seat tail delay analysis.

Access load forecasting uses load intensity forecasting based on Winter's seasonal exponential smoothing algorithm. Winter's linear and seasonal exponential smoothing methods utilize three equations, each of which is used to smooth the three components of the model: stationary, trend and seasonal, and both contain a relevant parameter. The three equations are as follows:

b_t＝γ(S_t+S_(t-1))+(1-γ)b_(t-1),0＜γ＜1 1-2

Among them, S _t refers to the exponential smoothing value of period t+1, x _t is the latest observation value, b _t refers to the correction term of the linear trend, ∝, β, and γ are smoothing coefficients, and the trend value is represented by adjacent The difference between the two smoothing values is expressed as: L is the length of the season; I _t is the seasonal correction coefficient of period t. For period t+m, its predicted value is expressed as:

F _t+m =(S _t +b _t m)I _t+mL ,0＜β＜1 (1-4)

Through repeated trials, the values of α, β and γ are determined to minimize the mean square error MSE to obtain the model. The mean square error is expressed as follows:

For the peak period and trough period of visit volume, the two time periods are predicted respectively, and their arrival rates are λ ₁ and λ ₂ respectively, which are used as the input parameters of tail delay analysis.

The tail delay analysis is based on the stochastic return network. Specifically, the different arrival rates at the peak and trough in the previous stage are modeled separately to explore the cloud service response time distribution under different access intensities. Calculating the response time distribution requires three steps. (1) Divide the total operation time t into n time slices; (2) In each time slice, obtain the corresponding response time distribution, and system failures should be taken into consideration in this process; (3) Obtain the overall response time distribution based on all time periods, obtain the tail delay under each quantile, and obtain its experience availability.

In terms of specific implementation, the tail delay analysis first uses the SPNP tool to calculate the stochastic return network model to obtain the initial state probability of the model in the stationary state, and then uses the label customer model to obtain the response time distribution in each initial state. After data Process and calculate to obtain the total response time distribution of user requests. The specific data processing and calculation process is as follows. The initial state of the system:

T＝[m,n,n',i,j,j',a,b,b',p,q,q'] (1)

m∈[0,…,M _s ],n∈[0,…,M _db ],n’∈[0,…,M _ca ],

i∈[0,…,N _s ],j∈[0,…,N _db ],j’∈[0,…,N _ca ],

a∈[0,…,N _s ],b∈[0,…,N _db ],b’∈[0,…,N _ca ],

p∈[0,…,N _s 4,q∈[0,…,N _db ],q’∈[0,…,N _ca ],

i+a+p＝N _s ,j+b+q＝N _db ,j′+b′+q′＝N _ca (2)

Among them, m is the current waiting queue length of the search layer, n is the current waiting queue length of the database layer, n' is the current waiting queue length of the cache layer, i is the current service queue length of the search layer, j is the current service queue length of the database layer, j' is the current service queue length of the cache layer, a is the number of currently failed connections in the search layer, b is the number of currently failed connections in the database layer, b' is the number of currently failed connections in the cache layer, p is the current idle and normal number of search layer The number of connections, q is the number of currently idle and normal connections in the database layer, q' is the number of currently idle and normal connections in the cache layer, M _s is the maximum length of the waiting queue in the search layer, M _db is the maximum length of the waiting queue in the database layer, N _s is the maximum number of connections supported by the search layer, N _db is the maximum number of connections supported by the database layer, N _ca is the maximum number of connections supported by the cache layer, π _x is the possibility that the system is in a stable state x, where Then the value of π _x is obtained through the stochastic return network model, and

At the same time, in the tag customer model, a non-empty state P _fin means that the tag customer's request has been processed. Therefore, the absorption state of the tag customer is defined as follows,

Among them, the reward function r _x t represents whether the absorption state is reached at time t;

By solving the reward function r _x t, we obtain the probability R _x t that the tag customer request is absorbed at time t under the initial state Therefore, calculate the probability Rt of completing the request processing process at time t, which is the cumulative distribution of system response time,

According to R(t), we know whether the delay exceeds τ at the current time slice, γ ^th percentile, and then directly use formula (6) to calculate the experience availability EA of the system,

in, Represents the high-tail delay, that is, at the γ ^th percentile, the delay is less than or equal to τ, that is, TL(γ)≤τ;/> Represents the low-tail delay, that is, at the γ ^th percentile, the delay is greater than τ, that is, TL(γ)>τ.

It also includes experimental verification steps: build an online search service experimental platform, use testing tools to send requests that obey specific distributions, simulate users to send requests to test the service capabilities of the search engine, and obtain the performance indicators of the service through testing as the theoretical values of the model. Then, the theoretical value of the response time distribution is obtained by combining the output results of the access load prediction method and the tail delay analysis method.

The characteristics and beneficial effects of the present invention are:

The present invention establishes a complete and accurate evaluation model for online cloud services, sets the real load intensity, hardware performance, and concurrency number to the corresponding arrival rate, failure rate, and number of connections, obtains the response time cumulative distribution function, and obtains the response time cumulative distribution function of the online cloud service. Tail latency characterizes its user experience, with fast speed and accurate results.

Picture description:

Figure 1 is an operation flow chart of the present invention;

Figure 2 shows the general architecture of online search services;

Figure 3 shows the stochastic return network model of online search service;

Figure 4 shows the tag customer model of online search services;

Figure 5 shows the cumulative distribution function of response time under changing the maximum number of connections;

In the figure, (a) request arrival rate = 100 (b) request arrival rate = 900.

Figure 6 Exponential smoothing prediction chart of weekly visits.

Detailed ways

The prediction method proposed by the present invention innovatively combines the access load prediction method and the tail delay analysis method on the basis of traditional service modeling analysis to achieve accurate prediction capabilities for the online cloud service response time distribution law. We applied this method to the online search service scenario, and added modeling analysis of node failure behavior and cache server acceleration behavior to the model, making it more consistent with the deployment and operation scenarios of current cloud services. Experimental results show that the present invention can accurately and effectively analyze the tail delay of online cloud services, providing a reference for ensuring user experience.

The present invention proposes an online cloud service performance prediction method based on time series prediction and stochastic return network, evaluates the response time of software applications under a specific hardware environment, and accurately depicts the tail delay of the 90th percentile, 99th percentile and other percentiles.

The cloud service response time distribution prediction method proposed by the present invention based on time series and random return network mainly consists of two modules: load intensity prediction and response time distribution prediction.

(1) Load intensity prediction based on Winter seasonal exponential smoothing algorithm

Winter's seasonal exponential smoothing algorithm can model time series with seasonal trends. There are seasonal patterns in user access to online cloud services. For example: taking 24 hours a day as a cycle, that is, seasonality. Most of the user visits every day are concentrated during the day, so the pressure on the server is mainly concentrated during the day, and it is often idle at night.

Winter's linear and seasonal exponential smoothing methods utilize three equations, each of which is used to smooth the three components of the model (stationary, trend and seasonal) and contains an associated parameter. The three equations are as follows:

b_t＝γ(S_t+S_(t-1))+(1-γ)b_(t-1),0＜γ＜1 (1-2)

Among them, S _t refers to the exponential smoothing value in period t+1, x _t is the latest observation value, b _t refers to the correction term of the linear trend, and the trend value is represented by the difference between two adjacent smoothing values, L is the length of the season; I is the seasonal correction coefficient.

For period t+m, its predicted value is expressed as:

F _t+ m＝(S _t +b _t m)I _t+mL ,0＜β＜1 (1-4)

Through trial and error method, the values of α, β and γ are determined to minimize the mean square error (MSE) to obtain the model. The expression of mean square error is as follows:

This model is used here, and the monthly visits to the NASA website are selected as the input data set to conduct time series analysis of the access arrival rate, thereby illustrating the application of this method. The prediction effect is shown in Figure 6:

Through experiments, the mean square error of this prediction is 0.19047. At this time, the values of α, β and γ are approximately equal to 0.27, 0.0 and 0.0. This method can predict the regular access rate of online cloud services (denoted as λ). As shown in the figure, the number of visits on working days from Monday to Friday is larger, which is regarded as the peak period; the number of visits on Saturdays and Sundays is smaller, it is regarded as a trough period, and the two time periods are predicted respectively, and their arrival rates are λ ₁ and λ ₂ respectively. The arrival rates are used as input parameters of the second module SRN model.

(2) Response time prediction based on Stochastic Reward Net (stochastic reward network)

Given the load arrival frequency prediction results of cloud computing services in different time periods, the present invention additionally provides a tail delay prediction method that combines a random return network and a tag customer model to achieve accurate modeling of user-perceived response time distribution. Since the SRN model only supports the Poisson distributed data flow that obeys a single variable as input, this invention will separately model different arrival rates at peaks and troughs in the previous stage to explore the distribution of cloud service response time under different access intensities. Calculating the response time distribution requires three steps: (1) Divide the total operation time t into n time slices. (2) In each time slice, the corresponding response time distribution is obtained, and system faults should be taken into consideration in this process. (3) Obtain the overall response time distribution based on all time periods, obtain the tail delay under each quantile, and thereby obtain its experience availability.

In terms of specific implementation, the present invention first uses tools such as SPNP (Stochastic Petri Net Package, Stochastic Petri Net Toolkit) to calculate the stochastic return network model to obtain the initial state probability of the model in a stationary state. Secondly, the label customer model is used to find the response time distribution in each initial state. After data processing and calculation, the total response time distribution of user requests is obtained. The specific data processing and calculation process are introduced below.

Assume that the initial state of the system is as follows:

T＝[m,n,n',i,j,j',a,b,b',p,q,q'] (1)

m∈[0,…,M _s ],n∈[0,…,M _db ],n’∈[0,…,M _ca ],

i∈[0,…,N _s ],j∈[0,…,N _db ],j’∈[0,…,N _ca ],

a∈[0,…,N _s ],b∈[0,…,N _db ],b’∈[0,…,N _ca ],

p∈[0,…,N _s ],q∈[0,…,N _db ],q’∈[0,…,N _ca ],

i+a+p＝N _s ,j+b+q＝N _db ,j′+b′+q′＝N _ca (2)

Among them, m is the current waiting queue length of the search layer, n is the current waiting queue length of the database layer, n' is the current waiting queue length of the cache layer, i is the current service queue length of the search layer, j is the current service queue length of the database layer, j' is the current service queue length of the cache layer, a is the number of currently failed connections in the search layer, b is the number of currently failed connections in the database layer, b' is the number of currently failed connections in the cache layer, p is the current idle and normal number of search layer The number of connections, q is the number of currently idle and normal connections in the database layer, q' is the number of currently idle and normal connections in the cache layer, M _s is the maximum length of the waiting queue in the search layer, M _db is the maximum length of the waiting queue in the database layer, N _s is the maximum number of connections supported by the search layer, N _db is the maximum number of connections supported by the database layer, and N _ca is the maximum number of connections supported by the cache layer.

Assume that π _x is the possibility that the system is in a stable state x, where Then the value of π _x can be obtained through the stochastic return network model, and

At the same time, in the tag customer model, a non-empty state P _fin means that the tag customer's request has been processed. Therefore, the absorptive state of the label customer is defined as follows,

Among them, the reward function r _x t represents whether the absorption state is reached at time t.

By solving the reward function r _x t, we can get the probability R _x t that the tag customer request is absorbed at time t under the initial state x. Therefore, the probability R(t) of completing the request processing process at time t can be calculated, that is, the cumulative distribution of system response time,

According to R(t), it is easy to know whether the delay exceeds τ at the current time slice, γ ^th percentile. Then, directly use formula (6) to calculate the experience availability EA of the system,

in, Represents the high-tail delay, that is, at the γ ^th percentile, the delay is less than or equal to τ, that is, TL(γ)≤τ; Represents the low-tail delay, that is, at the γ ^th percentile, the delay is greater than τ, that is, TL(γ)>τ.

The specific implementations, structures, features and functions provided by the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments.

1. The general structure of the online search service is as follows:

The present invention is modeled based on the general architecture of online search services, and its specific content is shown in Figure 2. The online search service mainly consists of three main components: web crawler component, index generator component and search component. The web crawler component crawls some accessible web pages from the Internet. The Index Builder component associates keywords found on these web pages with web page names to generate index information. Index information is stored in the database and is available for user search queries. The search component will accept the user's search request, support text analysis, and generate a list of search results by searching the index information stored in the database. During this process, the search component also weights each page in the entire list based on the information in the index and returns the final search results to the user.

The search component of the online search service mainly includes four layers: search layer, database layer and two load balancing layers. This invention does not consider these two load balancer layers because their processing time is much smaller than that of the search layer and the database layer. The database layer is divided into memory-based cache database and hard disk storage-based database. Depending on the number of users, each tier can deploy multiple server instances running in different virtual machines. Each instance, whether it's a search instance or a database instance, has a thread pool that accepts requests. When a search request arrives, it first waits in the search queue. The load balancer reads the search queue and distributes requests to specific instances of the search layer for text analysis. The search instance then allocates a connection thread from the thread pool to the request. If a connection thread is assigned a request, the connection thread will be occupied until the user receives the search results, and the connection thread will not be released. After the keywords are obtained through text analysis, the request will be further forwarded to the database instance. The database layer first determines whether to read data from the cache based on whether the cached data exists, whether the data has expired, etc. If valid data is found in the cache, the data is obtained from the cache and returned; otherwise, the database based on hard disk storage is queried. By setting the weight of accessing the cache and the weight of the database (such as 0.95 and 0.05), assign a thread to a specific database example to search all web pages that contain the keyword. Finally, the response editor in the search instance builds the list of web pages and sends the search results to the user.

2. The stochastic return network model of online search service is as follows:

This invention obtains the initial state probability of the system through a random return model. The random return model of the online search service is shown in Figure 3. Delayed transfer _Ta represents the request arrival process in the system. The states P _sw and P _dbw represent the waiting queues of the search layer and the database layer respectively. _Ta will only be triggered when the waiting queue length of the search layer is less than the longest waiting queue length. Once T _a (t _trans ) is triggered, a token will be added to the state P _sw (P _dbw ), that is, a request is added to the search layer (database layer) and is waiting to be served by the search layer (database layer) . State P _s and state P _db represent the number of idle resources in the search layer and database layer, that is, the number of idle connections. If a token is in state P _sw , and the number of tokens in state P _s is greater than or equal to 1, the token will be transferred from state P _sw to P _sp , and the number of tokens in state P _s will be reduced by one, that is, a request will be transferred from state P sw to P sp . The waiting queue is transferred to the service queue and consumes an idle connection. The status P _sp and the status P _dbp represent the service queues of the search layer and the database layer. The symbol # indicates that the actual trigger rate of the transfer is dependent. Therefore, the true firing rate of delayed transition T _sp is Kμ _s , where K is the number of tokens in state P _sp . After the token passes through the service queue of the search layer, it will be transferred to state P _trans , that is, the request has been served and left the search layer. Transient transfer t _trans represents the transfer of requests from the search layer to the database layer. Transient transfer to means that the request was rejected by the system. The transfer will only be triggered when the waiting queue at the database layer is full. If the waiting queue at the database layer is not full, the request will be transferred to state P _dbw . The database layer is similar to the search layer in that requests are transferred to the service queue only when resources are available. Finally, when the transfer T _dbp is triggered, a token will be removed from P _dbp , and the occupied search layer and database layer connection resources will also be released respectively, which means that a request has completed the entire search process and left system.

In addition, the stochastic reward network model proposed by the present invention can describe the fault occurrence and fault recovery behavior of the system. As shown in Figure 3, the state P _sd o _wn and the state P _dbd o _wn represent the number of failed connections in the search layer and database layer respectively. If a search instance fails, then the connections that the search instance is running in state P _sp , as well as the idle connections in state Ps, will fail. The transition T _sd represents the failure occurrence process of the search layer. The jagged lines in the figure represent the transfer of multiple tokens at once. Assume that in the system, I _s and I _db represent the number of instances of the search layer and database layer respectively. When an instance of the search layer fails, 1/Is tokens in state P _sp and 1/Is tokens in state P _s will be transferred to state P _sd o _wn at the same time, that is, all connections of the failed instance will be Transfer occurs and recovery is pending. Delayed transfer T _sr represents the recovery process of a failed instance. Once T _sr is triggered, all connections of the fault instance will transfer from state P _sd o _wn to state P _s , that is, the fault instance is restored to normal idle resources. The failure recovery process for the database layer is similar to that for the search layer. It is worth noting that since the occupied search layer and database layer resources are only released when the request leaves the database layer, when the database layer fails, in addition to transferring the failed database instance, the corresponding occupied search resources will also be released. Layer connection resources. The specific protection functions are shown in Table 1.

transfer protection function g1 #P _sw <M _s ? 1:0 g2 #P _dbw ≥M _db ? 1:0 g3 #P _dbw <M _db ? 1:0

Table 1

3. The tag customer model of the online search service is as follows:

The present invention uses the tag customer model to obtain the response time distribution in each initial state. The tag customer model of the online search service is shown in Figure 4. This model is similar to the random reward network model. It is worth noting that state P _csw represents a single tagged customer arrival queue with a queue length of 1. Assume that m, n, n', i, j, j', a, b, b', p, q, q' are states P _sw , P _dbw , P _caw , P _sp , P _dbp , P _cap , P respectively. _sdown , P _dbdown , P _cachedown , P _s , P _db , P _ca contain the number of tokens in the initial state, that is, the initial state of the system before the tag customer request arrives. In other words, when a tag customer request arrives, there are m(n',n) requests in the waiting queue and i(j',j) requests in the service queue at the search layer (cache layer, database layer) , a(b',b) connections fail, and p(q',q) idle connections can be occupied. The transient transition t _csw will only be triggered when the state P _sw is empty and P _s is not empty. Correspondingly, the instantaneous transfer t _cdbw will only be triggered when the state P _dbw is empty and P _db is not empty. The specific protection functions are shown in Table 2.

transfer protection function g1 #P _dbw ≥M _db ? 1:0 g2 #P _dbw <M _db ? 1:0 g3 #P _s >0 and #P _sw == 0? 1:0 g4 #P _db >0 and #P _dbw ==0?1:0 g5 #P _dbw ≥M _db ? 1:0 g6 #P _dbw <M _db ? 1:0

Table 2

4. Response time cumulative distribution function:

The present invention uses tools such as SPNP to calculate the random return network model and obtain the model value of the user request response time distribution. In addition, in terms of experimental verification, the present invention takes the Apache Solr search engine as an example to build an online search service experimental platform, uses testing tools such as Jmeter to send requests that obey a specific distribution (such as exponential distribution), and simulates users sending requests to search The service capability of the engine is tested, and the performance indicators of the service (such as service time, queuing time) are obtained through the test as the theoretical value of the model, and then the theoretical value of the response time distribution is obtained based on the output results of the model. Figure 5 illustrates the model values and experimental values of response time distribution by changing the request arrival rate to verify the correctness of the model. The user experience of the cloud service can be evaluated by analyzing the tail latency of each quantile of the online search service.

Claims (4)

1. A response time distribution prediction method for online cloud services, which is characterized by combining the access load prediction method and the tail delay analysis method to predict the response time distribution law of online cloud services, in which the access load prediction result is tail. Input parameters for delay analysis; access load forecast using load intensity forecast based on Winter's seasonal exponential smoothing algorithm, Winter's linear and seasonal exponential smoothing methods utilizing three equations, each of which is used to smooth the three components of the model Parts: stationary, trend and seasonal, and all contain a related parameter. The three equations are as follows: b_t＝γ(S_t+S_(t-1))+(1-γ)b_(t-1),0<γ<1 (1-2) Among them, S _t refers to the exponential smoothing value of period t+1, x _t is the latest observation value, b _t refers to the correction term of the linear trend, ∝, β, and γ are smoothing coefficients, and the trend value is represented by adjacent The difference between the two smoothing values is expressed as: L is the length of the season; I _t is the seasonal correction coefficient of period t. For period t+m, its predicted value is expressed as: F _t+m =(S _t +b _t m)I _t+mL ,0<β<1 (1-4) Through repeated trials, the values of α, β and γ are determined to minimize the mean square error MSE to obtain the model. The mean square error is expressed as follows: For the peak period and trough period of visit volume, the two time periods are predicted respectively, and their arrival rates are λ ₁ and λ ₂ respectively, which are used as the input parameters of tail delay analysis. 2. The response time distribution prediction method for online cloud services as claimed in claim 1, characterized in that the tail delay analysis is based on a stochastic return network. Specifically, different arrival rates at peaks and troughs in the previous stage are analyzed. Model separately to explore the response time distribution of cloud services under different access intensities. Calculating the response time distribution requires three steps: (1) Divide the total operation time t into n time slices; (2) In each time slice, obtain the corresponding response time distribution, and the system failure should also be taken into consideration in this process; (3) Obtain Based on the overall response time distribution in all time periods, the tail delay under each quantile is obtained to obtain its experience availability. 3. The response time distribution prediction method for online cloud services as claimed in claim 2, characterized in that, in specific implementation, the tail delay analysis first uses the SPNP tool to calculate the stochastic return network model to obtain the The initial state probability of the model, and then use the label customer model to find the response time distribution in each initial state. After data processing and calculation, the total response time distribution of user requests is obtained. The specific data processing and calculation process is as follows. The initial state of the system : T＝[m,n,n′,i,j,j′,a,b,b′,p,q,q′] (1) m∈[0,…,M _s ],n∈[0,…,M _db ],n′∈[0,…,M _ca ], i∈[0,…,N _s ],j∈[0,…,N _db ],j′∈[0,…,N _ca ], a∈[0,…,N _s ],b∈[0,…,N _db ],b′∈[0,…,N _ca ], p∈[0,…,N _s ], q∈[0,…,N _db ], q′∈[0,…,N _ca ], i+a+p＝N _s ,j+b+q＝N _db ,j′+b′+q′＝N _ca (2) Among them, m is the current waiting queue length of the search layer, n is the current waiting queue length of the database layer, n' is the current waiting queue length of the cache layer, i is the current service queue length of the search layer, j is the current service queue length of the database layer, j' is the current service queue length of the cache layer, a is the number of currently failed connections in the search layer, b is the number of currently failed connections in the database layer, b' is the number of currently failed connections in the cache layer, p is the current idle and normal number of search layer The number of connections, q is the number of currently idle and normal connections in the database layer, q' is the number of currently idle and normal connections in the cache layer, Ms is the maximum length of the waiting queue in the search layer, Mdb is the maximum length of the waiting queue in the database layer, Ns is the search The maximum number of connections supported by the layer, Ndb is the maximum number of connections supported by the database layer, Nca is the maximum number of connections supported by the cache layer, π _x is the possibility that the system is in a stable state x, where Then the value of π _x is obtained through the stochastic return network model, and At the same time, in the tag customer model, a non-empty state Pfin means that the tag customer's request has been processed. Therefore, the absorption state of the tag customer is defined as follows, Among them, the reward function r _x (t) indicates whether the absorption state is reached at time t; By solving the reward function r _x (t), we get the probability R _x (t) that the tag customer request is absorbed at time t under the initial state That is, the cumulative distribution of system response time, According to R(t), we know whether the delay exceeds τ at the current time slice, γ ^th percentile, and then directly use formula (6) to calculate the experience availability EA of the system, in, Represents the high-tail delay, that is, at the γ ^th percentile, the delay is less than or equal to τ, that is, TL(γ)≤τ; Represents the low-tail delay, that is, at the γ ^th percentile, the delay is greater than τ, that is, TL(γ)>τ. 4. The response time distribution prediction method for online cloud services as claimed in claim 1, further comprising an experimental verification step: building an online search service experimental platform, using testing tools to send requests that obey specific distributions, and simulating users to send Request to test the service capability of the search engine, obtain the performance indicators of the service through the test as the theoretical value of the model, and then combine the output results of the access load prediction method and the tail delay analysis method to obtain the theoretical value of the response time distribution.