CN110413657A - Average response time appraisal procedure towards seasonal form non-stationary concurrency - Google Patents
Average response time appraisal procedure towards seasonal form non-stationary concurrency Download PDFInfo
- Publication number
- CN110413657A CN110413657A CN201910624505.8A CN201910624505A CN110413657A CN 110413657 A CN110413657 A CN 110413657A CN 201910624505 A CN201910624505 A CN 201910624505A CN 110413657 A CN110413657 A CN 110413657A
- Authority
- CN
- China
- Prior art keywords
- concurrency
- stationary
- data
- form non
- response time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/21—Design, administration or maintenance of databases
- G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2474—Sequence data queries, e.g. querying versioned data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2477—Temporal data queries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The present invention provides a kind of average response time appraisal procedure towards seasonal form non-stationary concurrency, is related to field of cloud computer technology.This method is primarily based on the seasonal form non-stationary concurrency in the request Concurrency amount in auto-correlation coefficient method judgement cloud service system；Then the seasonal form non-stationary concurrency prediction model based on RNN-LSTM neural network is established, and carries out the prediction of seasonal form non-stationary concurrency；Establish the cloud service system average response time prediction model based on RBF, by the user's seasonal form non-stationary concurrency, cpu busy percentage, memory usage of prediction, these resource state informations for influencing cloud service average response time have been pre-processed later as input, export the average response time size for cloud service system.The present invention overcomes the deficiencies of traditional load balancing, improve the precision of prediction of seasonal form non-stationary concurrency, can make assessment response to service performance in time, cloud computing system is enable preferably to provide service for user.
Description
Technical field
The present invention relates to field of cloud computer technology more particularly to a kind of average responses towards seasonal form non-stationary concurrency
Time appraisal procedure.
Background technique
With the rapid development of Internet application more massive in today's society, cloud computing as a kind of new calculating and
Business model is also grown rapidly, so that user, which carries out the shared of various resources by internet, becomes more convenient.Cloud
Data center is the core of cloud service system, is mainly made of the various network hardwares with software equipment, not with user demand
Disconnected to improve, to cloud data center, higher requirements are also raised.Average response time is that the important performance in cloud service system refers to
One of mark, meanwhile, seasonal form non-stationary concurrency is widely present in cloud service system, not with seasonal form non-stationary concurrency
Disconnected variation, the service performance of cloud service system also will receive influence.
In cloud service system, seasonal form non-stationary concurrency is widely present, which peak occurs in the form of the period.
With the continuous variation of seasonal form non-stationary concurrency, the service performance of cloud service system will receive influence.Therefore season is faced
The continuous variation of type non-stationary concurrency, in order to ensure the reliability and stability of service performance, it will usually be adjusted using dynamic
The method of whole resource.But traditional load balancing needs the regular hour when formulating, especially when processing seasonal form is non-
The steady this load for occurring peak in the form of non-stationary of concurrency, easily lags behind real-time load state, and the lag loaded
The average response time that will lead to cloud service greatly increases, and the performance of cloud service system is also affected.
Summary of the invention
It is a kind of non-flat towards seasonal form the technical problem to be solved by the present invention is in view of the above shortcomings of the prior art, provide
The average response time of cloud service system is assessed in the average response time appraisal procedure of steady concurrency, realization.
In order to solve the above technical problems, the technical solution used in the present invention is: towards seasonal form non-stationary concurrency
Average response time appraisal procedure, comprising the following steps:
Step 1, based on auto-correlation coefficient method determine cloud service system in request Concurrency amount in seasonal form non-stationary simultaneously
Hair amount；
Step 1.1: the request Concurrency amount in cloud service system is extracted, using the mean value interpolation method in interpolation missing values to simultaneously
The data value of hair amount initial data missing is filled；
Request Concurrency amount in cloud service system is a two-dimensional array, the sets of numbers of the time and request initiated by request
At use [t, Ct] indicate, wherein at the time of t represents the request initiation that cloud service system monitors, CtRepresent t moment user visit
The number of request asked, then to the missing values C of the number of request of i moment user accessiIt is filled process, shown in following formula:
Wherein, CiIndicate the number of request of i moment user access, Ci-1Indicate the number of request of i-1 moment user access, Ci+1Table
Show the number of request of i+1 moment user access, Ci-1With Ci+1It is normal data；
Step 1.2: filled user's request Concurrency amount being divided into leveling style concurrency and seasonal form non-stationary type is concurrent
Amount；
The stationarity type concurrency is judged that the seasonal form non-stationary type is simultaneously according to the definition of leveling style time series
Hair amount is determined using auto-correlation coefficient method；
The seasonal form non-stationary concurrency is one kind of non-leveling style time series, and the judgement of its type is passed through certainly
Correlation analysis method carries out；When user requests to arrive, application this moment is monitored automatically every time cycle t cloud service system and is used
Family request Concurrency amount, and data are stored in database；T period concurrency observation contWith (t+k) period concurrent discharge observation
Value cont+kBetween degree of correlation, referred to as time delay be k concurrency auto-correlation coefficient rk；When concurrency auto-correlation coefficient
As the increase of time delay peak occurs with fixed frequency, and gradually it is intended to 0, i.e. rkIt is intended to judge that this is concurrent when 0
Amount is seasonal form non-stationary concurrency；
The time delay is the concurrency auto-correlation coefficient r of kkThe following formula of calculating shown in:
Wherein, Cov indicates covariance, and Var indicates variance；
Step 2 establishes the seasonal form non-stationary concurrency prediction model based on RNN-LSTM neural network, and carries out season
The prediction of type non-stationary concurrency, method particularly includes:
Step 2.1: the conversion of seasonal form non-stationary concurrency initial data；
Step 2.1.1: seasonal form non-stationary concurrency raw data format is converted, by the data extracted
Date is deleted, and concurrency data are only retained；
Step 2.1.2: to carried out format conversion seasonal form non-stationary concurrency data carry out Z-Score standardization and
Section scaling processing, and input and output data as RNN-LSTM neural network；
(a) Z-Score standardization is carried out to seasonal form non-stationary concurrency data
The concurrency data for having rejected time dimension are denoted as X={ x1, x2..., xn, to concurrency sequence X={ x1,
x2..., xnCarry out Z-Score standardization after, the new sequences y of generation1, y2..., yn, wherein
The new sequences y generated1, y2..., ynMean value be 0, and variance be 1；
(b) section scaling is carried out to seasonal form non-stationary concurrency data
The value interval of feature in concurrency data is transformed into [0,1] range, normalization formula is as follows:
Wherein, X indicates concurrency data, and Min indicates the minimum value in concurrency data, and Max is indicated in concurrency data
Maximum value, Y indicates the data that have handled；
Step 2.2: generating model training data set；Determine training set length of window and training set and test set data
Size；The determination for being determined as the seasonal form non-stationary concurrency period of the training set length of window, by analyzing auto-correlation system
Number figure obtains, for clearly predicting the time interval of next concurrency；
The determination and adjustment of step 2.3:RNN-LSTM Artificial Neural Network Structures；The RNN-LSTM neural network model
Including memory block and each input gate for handling data, out gate and forget door；
Step 2.3.1: each memory block is expressed as a Cell；Using memory block as the basic of network concealed layer
Unit, it includes one or more memory cells and a pair of adaptive multiplication door control unit, the door control unit is by concurrency
Input and the output of concurrency be connected to all units in the block；The core of each memory cell has one to be referred to as
Constant error conveyer belt, that is, CEC circulation is from connecting linear unit, the referred to as location mode when it is activated；In not new input
Or when error signal, the local error of CEC is remained unchanged, and is neither increased also unattenuated；CEC by output and input respectively by
To the preceding protection to excitation and backward error；When the door is closed, uncorrelated input and noise do not enter unit, and location mode is not
The rest part of network can be interfered；
Step 2.3.2: input gate and out gate are calculated；The input gate and out gate be it is a kind of allow information selection pass through
Method, acts on the input and output of parameter, to control the data of input and output；Outputting and inputting door includes one
Sigmoid neural network and a pointwise multiplication operation；Numerical value between Sigmoid layers of output 0 to 1, description network are every
How many concurrency of a part passes through, and 0 represents " disagreeing any amount to pass through ", and 1 represents " any amount is allowed to pass through "；
Step 2.3.3:: door is forgotten in calculating；It introduces and forgets that Men Ze is that gradually resetting corresponds to the location mode slowly to decline；
Forget that the effect of door is similar with input output gate, unlike, when the training stage starts, forget that the activation of door is 1, it is entire single
LSTM unit of the behavior of member just as a standard；
After step 2.4:RNN-LSTM Establishment of Neural Model, the seasonal form non-stationary concurrency data handled well are passed
Enter model, the epoch of suitable training dataset and test data set and model training is set, determines the meter of model error
Calculation mode carries out the prediction of seasonal form non-stationary concurrency；
Step 3 establishes the cloud service system average response time prediction model based on RBF, by user's seasonal form of prediction
These resource state informations for influencing cloud service average response time of non-stationary concurrency, cpu busy percentage, memory usage are located in advance
As input after having managed, the average response time size for cloud service system is exported, method particularly includes:
Step 3.1: data collection and pretreatment；The data of collection four seed type of cloud service system, including history concurrency,
Average response time in CPU utilization rate, memory usage and each service；And to the data of collection before establishing model
It is standardized using section scaling and Z-Score standardized method；
Step 3.2: establishing the cloud service system average response time prediction model based on RBF network；
Step 3.2.1: determine the extension constant of RBF network, the center of radial basis function and hidden layer to output layer
Three network parameters of weight；
(1) extension constant is determined
Extension constant is also known as width vector, represents hidden layer neuron to the induction range of input information；Width is got over
Small induction range is smaller, on the contrary, the bigger induction range of width is bigger；The initialization width vector of RBF network isWherein,
Wherein, dji′For the i-th ' a value of j-th of width vector, i '=1,2 ..., n, n is RBF neural input layer
Unit number, dfFor width adjusting coefficient, cji′Indicate the Center Parameter of radial basis function, N is given constant, and r is that circulation becomes
Amount；
(2) center of radial basis function is determined
The Center Parameter of radial basis function is for completing the mapping of input layer to hidden layer；Each of hidden layer nerve
Member has a radial basis function center, which is expressed as Cj=[Cj1, Cj2... Cjn]T, wherein
Wherein, q is the network concealed node layer number of RBF, miniWith maxiIt is all defeated to respectively indicate ith feature in training set
Enter the minimum value and maximum value of information；
(3) determine hidden layer to output layer weight
Weight W of the hidden layer to output layers=[ws1, ws2... wsp]T, s=1,2, wherein
Wherein, maxsWith minsRespectively indicate all output informations i.e. average response in s-th of output neuron in training set
The maxima and minima of time；
Step 3.2.2: the output Z of network concealed j-th of the neuron of layer of RBF is calculatedjIt is real with the output of output layer neuron
Now construct complete RBF neural；
Wherein, the output calculation formula of hidden layer are as follows:
The output calculation formula of output layer neuron are as follows:
O=[o1, o2]T (11)
Wherein, O is the output of output layer neuron；
Step 3.3: the cloud service system average response time prediction model based on RBF network of foundation is trained,
Obtain optimal average time response prediction model；
Step 3.3.1: by the history of the different moments of Web same on the virtual machine in cloud service system application respectively serviced
The season in memory usage, CPU usage, average response time, and each service of prediction is inscribed when concurrency and correspondence
Nodal pattern non-stationary concurrency data reach the input node of RBF network；
Step 3.3.2: using the center of the radial basis function of K-Means method initialization RBF network, stochastic gradient is used
Algorithm adaptively adjusts the weight iteration of three network parameters in the RBF network of building；
Step 3.3.3: three parameters of initialization RBF network: extension constant Dji, hidden layer radial basis function center
DjiWith the connection weight W of hidden layer to output layerkj；
Step 3.3.4: setting hidden layer interstitial content, Studying factors, model training error precision ε and maximum frequency of training
MT；
Step 3.3.5: loop initialization the number of iterations cycle=1；
Step 3.3.6: based on the stochastic gradient descent method training big parameter of RBF network three, optimize the cloud clothes based on RBF network
Business Mean Time of Systemic Response prediction model；
Step 3.3.7: the root-mean-square error (Root Mean Square Error, i.e. RMSE) of computation model training, such as
Fruit RMSE >=ε, i.e. RMSE are greater than a very small value, then cycle=cycle+1, no to then follow the steps 3.3.9；
Step 3.3.8: judging whether cycle is less than maximum frequency of training MT, if it is less, the 3.3.6 that gos to step,
It is no to then follow the steps 3.3.9；
Step 3.3.9: end loop obtains optimal average time response prediction model；
Step 3.4: step 3.1 is collected and is input to optimal average time response prediction model with pretreated data,
The average response time for the cloud service system predicted.
The beneficial effects of adopting the technical scheme are that one kind provided by the invention is towards seasonal form non-stationary
The average response time appraisal procedure of concurrency, using RNN-LSTM neural network to the seasonal form non-stationary concurrency handled
It is predicted, and as one of data input, is input in RBF network model and predicts average response time, to cloud service
The performance of system is assessed.The present invention overcomes the deficiencies of traditional load balancing, improve seasonal form non-stationary simultaneously
The precision of prediction of hair amount improves the assessment accuracy to average response time, can make assessment response to service performance in time,
Cloud computing system is set preferably to provide service for user.
Detailed description of the invention
Fig. 1 is the average response time appraisal procedure provided in an embodiment of the present invention towards seasonal form non-stationary concurrency
Flow chart；
Fig. 2 is the experiment topological diagram that embodiment of the invention provides；
Fig. 3 is the concurrency initiation sequence figure that embodiment of the invention provides；
Fig. 4 be embodiment of the invention provide using ARIMA method and RNN-LSTM neural network of the invention to portion
Divide the prediction result comparison diagram of concurrency；
Fig. 5 be embodiment of the invention provide using BP algorithm and improved RBF network model of the invention to average sound
The comparison diagram of prediction result between seasonable.
Specific embodiment
With reference to the accompanying drawings and examples, specific embodiments of the present invention will be described in further detail.Implement below
Example is not intended to limit the scope of the invention for illustrating the present invention.
The present embodiment is by taking the instamatic cloud service system that certain laboratory is built as an example, using of the invention towards seasonal form
The average response time appraisal procedure of non-stationary concurrency assesses the average response time of the cloud service system.
Average response time appraisal procedure towards seasonal form non-stationary concurrency, as shown in Figure 1, comprising the following steps:
Step 1, based on auto-correlation coefficient method determine cloud service system in request Concurrency amount in seasonal form non-stationary simultaneously
Hair amount；
Step 1.1: the request Concurrency amount in cloud service system is extracted, using the mean value interpolation method in interpolation missing values to simultaneously
The data value of hair amount initial data missing is filled；
Request Concurrency amount in cloud service system is a two-dimensional array, the sets of numbers of the time and request initiated by request
At use [t, Ct] indicate, wherein at the time of t represents the request initiation that cloud service system monitors, CtRepresent t moment user visit
The number of request asked, then to the missing values C of the number of request of i moment user accessiIt is filled process, shown in following formula:
Wherein, CiIndicate the number of request of i moment user access, Ci-1Indicate the number of request of i-1 moment user access, Ci+1Table
Show the number of request of i+1 moment user access, Ci-1With Ci+1It is normal data；
Step 1.2: filled user's request Concurrency amount being divided into leveling style concurrency and seasonal form non-stationary type is concurrent
Amount；
The stationarity type concurrency is judged that the seasonal form non-stationary type is simultaneously according to the definition of leveling style time series
Hair amount is determined using auto-correlation coefficient method；
The seasonal form non-stationary concurrency is one kind of non-leveling style time series, and the judgement of its type is passed through certainly
Correlation analysis method carries out；When user requests to arrive, application this moment is monitored automatically every time cycle t cloud service system and is used
Family request Concurrency amount, and data are stored in database；T period concurrency observation contWith (t+k) period concurrent discharge observation
Value cont+kBetween degree of correlation, referred to as time delay be k concurrency auto-correlation coefficient rk；When concurrency auto-correlation coefficient
As the increase of time delay peak occurs with fixed frequency, and gradually it is intended to 0, i.e. rkIt is intended to judge that this is concurrent when 0
Amount is seasonal form non-stationary concurrency；
The time delay is the concurrency auto-correlation coefficient r of kkThe following formula of calculating shown in:
Wherein, Cov indicates covariance, and Var indicates variance；
Step 2 establishes the seasonal form non-stationary concurrency prediction model based on RNN-LSTM neural network, and carries out season
The prediction of type non-stationary concurrency, method particularly includes:
Step 2.1: the conversion of seasonal form non-stationary concurrency initial data；
Step 2.1.1: seasonal form non-stationary concurrency raw data format is converted, by the data extracted
Date is deleted, and concurrency data are only retained；In order to improve the quality of data, convenient incoming model is predicted, by extraction
Seasonal form non-stationary concurrency initial data carries out certain conversion.It is the conversion of format first, the seasonal form directly extracted is non-
Steady concurrency data are for convenience directly predicted data, the data that will be extracted with determining time dimension
Date deleted, only retain concurrency data.
Step 2.1.2: to carried out format conversion seasonal form non-stationary concurrency data carry out Z-Score standardization and
Section scaling processing, and input and output data as RNN-LSTM neural network；In view of LSTM method for the big of data
It is small more sensitive, therefore input and output data to algorithm is also needed to handle accordingly, the specific steps are as follows:
(a) Z-Score standardization (Z-Score Standardization) is carried out to seasonal form non-stationary concurrency data
Z-Score standardized method uses the average value and standard deviation of initial data, carries out the standardization of data, it is suitable
The unknown situation of maxima and minima for attribute to be treated, or have the case where part Outlier Data.It will reject
The concurrency data of complete time dimension are denoted as X={ x1, x2..., xn, to concurrency sequence X={ x1, x2..., xnCarry out Z-
After Score standardization, the new sequences y of generation1, y2..., yn, wherein
The new sequences y generated1, y2..., ynMean value be 0, and variance be 1；
(b) section scaling (Min-Max Scaling) is carried out to seasonal form non-stationary concurrency data
Section scaling is one kind of data normalization, it can zoom in and out the mode that initial data linearizes, will
The value interval of feature is transformed into [0,1] range in concurrency data, and normalization formula is as follows:
Wherein, X indicates concurrency data, and Min indicates the minimum value in concurrency data, and Max is indicated in concurrency data
Maximum value, Y indicates the data that have handled；
Step 2.2: generating model training data set；Determine training set length of window and training set and test set data
Size；The determination for being determined as the seasonal form non-stationary concurrency period of the training set length of window, by analyzing auto-correlation system
Number figure obtains, for clearly predicting the time interval of next concurrency；
The determination and adjustment of step 2.3:RNN-LSTM Artificial Neural Network Structures；The RNN-LSTM neural network model
Including memory block and each input gate for handling data, out gate and forget door；
Step 2.3.1: each memory block is expressed as a Cell；Using memory block as the basic of network concealed layer
Unit, it includes one or more memory cells and a pair of adaptive multiplication door control unit, the door control unit is by concurrency
Input and the output of concurrency be connected to all units in the block；The core of each memory cell has one to be referred to as
Constant error conveyer belt, that is, CEC circulation is from connecting linear unit, the referred to as location mode when it is activated；In not new input
Or when error signal, the local error of CEC is remained unchanged, and is neither increased also unattenuated；CEC by output and input respectively by
To the preceding protection to excitation and backward error；When the door is closed, uncorrelated input and noise do not enter unit, and location mode is not
The rest part of network can be interfered；
Step 2.3.2: input gate and out gate are calculated；The input gate and out gate be it is a kind of allow information selection pass through
Method, acts on the input and output of parameter, to control the data of input and output；Outputting and inputting door includes one
Sigmoid neural network and a pointwise multiplication operation；Numerical value between Sigmoid layers of output 0 to 1, description network are every
How many concurrency of a part passes through, and 0 represents " disagreeing any amount to pass through ", and 1 represents " any amount is allowed to pass through "；
In the present embodiment, input gate al tCalculating it is as follows:
Wherein,
bl t=f (al t)
bl t=f (al t) it is activation primitive, the effect of activation primitive is to add non-linear factor for function, improves nerve net
The ability to express of network.
Out gate aw tBe expressed as follows:
Wherein,
bw t=f (aw t)
Step 2.3.3:: door is forgotten in calculating；It introduces and forgets that Men Ze is that gradually resetting corresponds to the location mode slowly to decline；
Forget that the effect of door is similar with input output gate, unlike, when the training stage starts, forget that the activation of door is 1, it is entire single
LSTM unit of the behavior of member just as a standard；
In the present embodiment, forget that being expressed as follows for door is shown:
Wherein,
bφ t=f (aφ t)
Last output:
bc t=bw th(sc t)
After step 2.4:RNN-LSTM Establishment of Neural Model, the seasonal form non-stationary concurrency data handled well are passed
Enter model, the epoch of suitable training dataset and test data set and model training is set, determines the meter of model error
Calculation mode carries out the prediction of seasonal form non-stationary concurrency；
Step 3 establishes the cloud service system average response time prediction model based on RBF, by user's seasonal form of prediction
These resource state informations for influencing cloud service average response time of non-stationary concurrency, cpu busy percentage, memory usage are located in advance
As input after having managed, the average response time size for cloud service system is exported, method particularly includes:
Step 3.1: data collection and pretreatment；The data of collection four seed type of cloud service system, including history concurrency,
Average response time in CPU utilization rate, memory usage and each service；And to the data of collection before establishing model
It is standardized using section scaling and Z-Score standardized method；
Step 3.2: establishing the cloud service system average response time based on RBF network (i.e. radial primary function network) and predict
Model；
Step 3.2.1: determine the extension constant of RBF network, the center of radial basis function and hidden layer to output layer
Three network parameters of weight；
(1) extension constant is determined
Extension constant is also known as width vector, represents hidden layer neuron to the induction range of input information；Width is got over
Small induction range is smaller, on the contrary, the bigger induction range of width is bigger；The initialization width vector of RBF network isWherein,
Wherein, dji′For the i-th ' a value of j-th of width vector, i '=1,2 ..., n, n is RBF neural input layer
Unit number, dfFor width adjusting coefficient, cji′Indicate the Center Parameter of radial basis function, N is given constant, and r is that circulation becomes
Amount；
(2) center of radial basis function is determined
The Center Parameter of radial basis function is for completing the mapping of input layer to hidden layer；Each of hidden layer nerve
Member has a radial basis function center, which is expressed as Cj=[Cj1, Cj2... Cjn]T, wherein
Wherein, q is the network concealed node layer number of RBF, miniWith maxiIt is all defeated to respectively indicate ith feature in training set
Enter the minimum value and maximum value of information；
(3) determine hidden layer to output layer weight
Weight W of the hidden layer to output layers=[ws1, ws2... wsp]T, s=1,2, wherein
Wherein, maxsWith minsRespectively indicate all output informations i.e. average response in s-th of output neuron in training set
The maxima and minima of time；
Step 3.2.2: the output Z of network concealed j-th of the neuron of layer of RBF is calculatedjIt is real with the output of output layer neuron
Now construct complete RBF neural；
Wherein, the output calculation formula of hidden layer are as follows:
The output calculation formula of output layer neuron are as follows:
O=[o1, o2]T (11)
Wherein, O is the output of output layer neuron；
Step 3.3: the cloud service system average response time prediction model based on RBF network of foundation is trained,
Obtain optimal average time response prediction model；
Step 3.3.1: by the history of the different moments of Web same on the virtual machine in cloud service system application respectively serviced
The season in memory usage, CPU usage, average response time, and each service of prediction is inscribed when concurrency and correspondence
Nodal pattern non-stationary concurrency data reach the input node of RBF network；
Step 3.3.2: using the center of the radial basis function of K-Means method initialization RBF network, stochastic gradient is used
Algorithm adaptively adjusts the weight iteration of three network parameters in the RBF network of building；
Step 3.3.3: three parameters of initialization RBF network: extension constant Dji, hidden layer radial basis function center
DjiWith the connection weight W of hidden layer to output layerkj；
Step 3.3.4: setting hidden layer interstitial content, Studying factors, model training error precision ε and maximum frequency of training
MT；
Step 3.3.5: loop initialization the number of iterations cycle=1；
Step 3.3.6: based on the stochastic gradient descent method training big parameter of RBF network three, optimize the cloud clothes based on RBF network
Business Mean Time of Systemic Response prediction model；
Step 3.3.7: the root-mean-square error (Root Mean Square Error, i.e. RMSE) of computation model training, such as
Fruit RMSE >=ε, i.e. RMSE are greater than a very small value, then cycle=cycle+1, no to then follow the steps 3.3.9；
Step 3.3.8: judging whether cycle is less than maximum frequency of training MT, if it is less, the 3.3.6 that gos to step,
It is no to then follow the steps 3.3.9；
Step 3.3.9: end loop obtains optimal average time response prediction model；
Step 3.4: step 3.1 is collected and is input to optimal average time response prediction model with pretreated data,
The average response time for the cloud service system predicted.
All data in the present embodiment are collected in instamatic cloud service system, are that a collection is registered, logged in, ordering
The intelligent seat reservation system of a variety of cloud services such as ticket.Server used in the instamatic cloud service system is
SugonI620-G20, wherein the IP of two-server is respectively 202.199.6.137,202.199.6.106.Server configuration
As shown in table 1.
1 server configuration table of table
It all include more virtual machines in every physical machine of the cloud service system.Meanwhile using virtual machine manager
(Kernel-based Virtual Machine, i.e. KVM) installs virtual machine on the server, completes the virtualization of environment.
The present embodiment is virtual by this tool management of virt-manager KVM other than using KVM to create virtual machine
Environment.Before creating virtual machine, CentOS mirror image is put into server, the peace of virtual machine is carried out using virt-manager
The mirror image can be called directly when dress.
The virtual machine configuration situation that the present embodiment uses is as shown in table 2, and using the virtual machine of isomorphism, i.e. each VM's
Configuring condition is all identical.CPU core calculation in the virtual machine used is 4 core, the memory size of 2G, the hard-disk capacity of 20G and
The operating system of CentOS.
2 virtual machine configuration table of table
| Performance indicator | Performance parameter |
| CPU core calculation | 4 cores |
| Memory size | 2G |
| Hard-disk capacity | 20GB |
| Operating system | CentOS |
After having built the environment of the bottom, needs to install software in the environment for the collection of virtual-machine data and divide
Analysis.In the present embodiment there are two types of used system performance testing softwares: first set is Collectd+InfluxDB+
Grafana, wherein Collectd is used for the acquisition of virtual machine performance data, and InfluxDB is deposited for virtual machine performance data
Storage, Grafana are used for the display of virtual machine performance data；Another kind is directly to carry out performance to system using LoadRunner
Detection.When using LoadRunner, there are four parts altogether, be respectively creation script, design simulation scene, Run-time scenario with
And analysis result.
Finally, installing VNC-Viewer on the computer of laboratory, it is used to remotely connect service in window client
Device.The virtual environment built elsewhere can be controlled in laboratory.For the experimental situation, topological diagram such as Fig. 2 institute is tested
Show.
During the collection of data, the creation for carrying out script is first had to.When carrying out recording script using LoadRunner,
By to seasonal form non-stationary concurrent request system and webpage record.After opening LoadRunner, first
Creation or edit script, i.e. opening VuGen start page are needed, new Web script is created, starts to record.Main needle when recording
Seasonal form non-stationary concurrent request is recorded.After having recorded, it can be checked by tree view or script view,
Also it can specify certain affairs to be operated.After having created script, next need to build load testing environment.It utilizes
Controller carries out load testing, first has to creation scene, and script created is loaded into the test environment of load
In.Load Generator adds the computer for generating and loading by running Vuser in the application, and can be same
One time used multiple Load Generator, ran multiple Vuser in each Load Generator.Controller exists
Load Generator can be independently connected to when Run-time scenario.After scene has been run, the extraction of data can be carried out.Its
In, part concurrency data such as Fig. 3 is shown, the present embodiment choose user concurrent amount data be using 1min as time interval,
The concurrency data that 130 historical users access instamatic system simultaneously are training set.
After the historical user's concurrency data for analyzing acquisition, need to pre-process data, it is complete that data, which are supplemented,
Then whole data set judges the trend type of user concurrent amount data using auto-correlation coefficient method, and takes the period
The mode of difference carries out tranquilization processing to user concurrent amount data.Later, using RNN-LSTM prediction model of the invention into
Row prediction, and compared with truthful data, verify the validity of model.After having been predicted using RNN-LSTM and ARIMA
It is as shown in table 3 to compare data.
Table 3 predicts error comparison
| Algorithm | Average relative error | Averaged Square Error of Multivariate | Mean absolute error |
| ARIMA | 0.0586 | 6.7584 | 3.5864 |
| RNN-LSTM | 0.0467 | 4.3654 | 2.0157 |
By comparative analysis it is found that the prediction accuracy of RNN-LSTM prediction model of the invention is higher than ARIMA.Fig. 4
Prediction for ARIMA and RNN-LSTM to part concurrency.As shown in Table 3, the average relative error of RNN-LSTM, average square
Error and mean absolute error are respectively 0.0586,6.7584 and 3.5864, and the three classes error of ARIMA is respectively
0.0467,4.3654 and 2.0157, therefore show in seasonal form non-stationary concurrency forecasting problem, RNN-LSTM tool
There is higher precision of prediction.
After user's history concurrency data have been predicted, while the every average response time for monitoring influence cloud service refers to
Mark, and data prediction is carried out to each index, and then the average response time for establishing the concurrency prediction based on RBF predicts mould
Type.Meanwhile the present embodiment obtains the average response time prediction result that the method for the present invention obtains with by BP neural network algorithm
It is compared to average response time, obtains comparing result as shown in Figure 5.
It is analyzed by comparing result it is found that the prediction of improved RBF average response time prediction model of the invention is accurate
Property is higher than BP algorithm.As shown in table 4, the average relative error of RBF, Averaged Square Error of Multivariate and mean absolute error are respectively
0.0125,1.5642 and 1.2896, and the three classes error of BP is respectively 0.0425,5.3869 and 3.9865, shows RBF
There is better approximation capability for non-linear relation combination, therefore shows the cloud service existing for web services user concurrent amount
In average response time forecasting problem, the method for the present invention precision of prediction with higher and lower prediction error.
The comparison of 4 average response time error of table
| Algorithm | Average relative error | Averaged Square Error of Multivariate | Mean absolute error |
| BP | 0.0425 | 5.3869 | 3.9865 |
| RBF | 0.0125 | 1.5642 | 1.2896 |
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify to technical solution documented by previous embodiment, or some or all of the technical features are equal
Replacement；And these are modified or replaceed, model defined by the claims in the present invention that it does not separate the essence of the corresponding technical solution
It encloses.
Claims (8)
1. a kind of average response time appraisal procedure towards seasonal form non-stationary concurrency, it is characterised in that: including following step
It is rapid:
Step 1, based on auto-correlation coefficient method determine cloud service system in request Concurrency amount in seasonal form non-stationary concurrency；
Step 1.1: the request Concurrency amount in cloud service system is extracted, using the mean value interpolation method in interpolation missing values to concurrency
The data value of initial data missing is filled；
Step 1.2: filled user's request Concurrency amount is divided into leveling style concurrency and seasonal form non-stationary type concurrency；
Step 2 establishes the seasonal form non-stationary concurrency prediction model based on RNN-LSTM neural network, and it is non-to carry out seasonal form
Steady concurrency prediction, method particularly includes:
Step 2.1: the conversion of seasonal form non-stationary concurrency initial data；
Step 2.1.1: seasonal form non-stationary concurrency raw data format is converted, by the date of the data extracted
It is deleted, only retains concurrency data；
Step 2.1.2: Z-Score standardization and section are carried out to the seasonal form non-stationary concurrency data for having carried out format conversion
Scaling processing, and input and output data as RNN-LSTM neural network；
Step 2.2: generating model training data set；Determine the big of training set length of window and training set and test set data
It is small；The determination for being determined as the seasonal form non-stationary concurrency period of the training set length of window, by analyzing auto-correlation coefficient
Figure obtains, for clearly predicting the time interval of next concurrency；
The determination and adjustment of step 2.3:RNN-LSTM Artificial Neural Network Structures；The RNN-LSTM neural network model includes
Memory block and each processing input gate of data, out gate and forget door；
After step 2.4:RNN-LSTM Establishment of Neural Model, the seasonal form non-stationary concurrency data handled well are passed to mould
Type is arranged the epoch of suitable training dataset and test data set and model training, determines the calculating side of model error
Formula carries out the prediction of seasonal form non-stationary concurrency；
Step 3 establishes the cloud service system average response time prediction model based on RBF, and user's seasonal form of prediction is non-flat
These resource state informations for influencing cloud service average response time of steady concurrency, cpu busy percentage, memory usage have pre-processed
Later as input, the average response time size for cloud service system is exported, method particularly includes:
Step 3.1: data collection and pretreatment；Collect the data of four seed type of cloud service system, including history concurrency, CPU
Average response time in utilization rate, memory usage and each service；And the data of collection are used before establishing model
Section scaling and Z-Score standardized method are standardized；
Step 3.2: establishing the cloud service system average response time prediction model based on RBF network；
Step 3.2.1: determine RBF network extension constant, radial basis function center and hidden layer to output layer weight
Three network parameters；
Step 3.2.2: the output Z of network concealed j-th of the neuron of layer of RBF is calculatedjWith the output of output layer neuron, structure is realized
Build complete RBF neural；
Step 3.3: the cloud service system average response time prediction model based on RBF network of foundation being trained, is obtained
Optimal average time response prediction model；
Step 3.4: step 3.1 being collected and is input to optimal average time response prediction model with pretreated data, is obtained
The average response time of the cloud service system of prediction.
2. the average response time appraisal procedure according to claim 1 towards seasonal form non-stationary concurrency, feature
It is: step 1.1 method particularly includes:
Request Concurrency amount in cloud service system is a two-dimensional array, and the quantity of the time and request initiated by request form,
Use [t, Ct] indicate, wherein at the time of t represents the request initiation that cloud service system monitors, CtRepresent t moment user access
Number of request, then to i moment user access number of request missing values CiIt is filled process, shown in following formula:
Wherein, CiIndicate the number of request of i moment user access, Ci-1Indicate the number of request of i-1 moment user access, Ci+1Indicate i+1
The number of request of moment user access, Ci-1With Ci+1It is normal data.
3. the average response time appraisal procedure according to claim 1 towards seasonal form non-stationary concurrency, feature
It is: step 1.2 method particularly includes:
The stationarity type concurrency is defined according to leveling style time series to be judged, the seasonal form non-stationary type concurrency
Determined using auto-correlation coefficient method；
The seasonal form non-stationary concurrency is one kind of non-leveling style time series, passes through auto-correlation for the judgement of its type
Coefficient analysis method carries out；When user requests to arrive, monitors every time cycle t cloud service system and asked this moment using user automatically
Concurrency is sought, and data are stored in database；T period concurrency observation contWith (t+k) period concurrency observation
cont+kBetween degree of correlation, referred to as time delay be k concurrency auto-correlation coefficient rk；When concurrency auto-correlation coefficient with
The increase of time delay there is peak with fixed frequency, and be gradually intended to 0, i.e. rkIt is intended to judge the concurrency when 0
For seasonal form non-stationary concurrency；
The time delay is the concurrency auto-correlation coefficient r of kkThe following formula of calculating shown in:
Wherein, Coy indicates covariance, and Var indicates variance.
4. the average response time appraisal procedure according to claim 1 towards seasonal form non-stationary concurrency, feature
Be: it is standardized to carry out Z-Score to seasonal form non-stationary concurrency data described in step 2.1.2 method particularly includes:
The concurrency data for having rejected time dimension are denoted as X={ x1, x2..., xn, to concurrency sequence X={ x1,
x2..., xnCarry out Z-Score standardization after, the new sequences y of generation1, y2..., yn, wherein
The new sequences y generated1, y2..., ynMean value be 0, and variance be 1；
It is described that section scaling is carried out to seasonal form non-stationary concurrency data method particularly includes:
The value interval of feature in concurrency data is transformed into [0,1] range, normalization formula is as follows:
Wherein, X indicates concurrency data, and Min indicates the minimum value in concurrency data, and Max is indicated in concurrency data most
Big value, Y indicate the data handled.
5. the average response time appraisal procedure according to claim 1 towards seasonal form non-stationary concurrency, feature
It is: the step 2.3 method particularly includes:
Step 2.3.1: each memory block is expressed as a Cell；Using memory block as the substantially single of network concealed layer
Member, it includes one or more memory cells and a pair of adaptive multiplication door control unit, the door control unit is by concurrency
Input and the output of concurrency are connected to all units in the block；The core of each memory cell has one to be referred to as perseverance
Determine the error conveyer belt i.e. circulation of CEC from connecting linear unit, the referred to as location mode when it is activated；In not new input or
When error signal, the local error of CEC is remained unchanged, and is neither increased also unattenuated；CEC by output and input respectively by
The protection of forward direction excitation and backward error；When the door is closed, uncorrelated input and noise do not enter unit, and location mode will not
Interfere the rest part of network；
Step 2.3.2: input gate and out gate are calculated；The input gate and out gate are a kind of sides for allowing information to select to pass through
Method, acts on the input and output of parameter, to control the data of input and output；Outputting and inputting door includes one
Sigmoid neural network and a pointwise multiplication operation；Numerical value between Sigmoid layers of output 0 to 1, description network are every
How many concurrency of a part passes through, and 0 represents " disagreeing any amount to pass through ", and 1 represents " any amount is allowed to pass through "；
Step 2.3.3:: door is forgotten in calculating；It introduces and forgets that Men Ze is that gradually resetting corresponds to the location mode slowly to decline；Forget
The effect of door is similar with input output gate, unlike, when the training stage starts, forget that the activation of door is 1, entire unit
LSTM unit of the behavior just as a standard.
6. the average response time appraisal procedure according to claim 1 towards seasonal form non-stationary concurrency, feature
It is: the step 3.2.1's method particularly includes:
(1) extension constant is determined
Extension constant is also known as width vector, represents hidden layer neuron to the induction range of input information；The smaller sense of width
Answer range smaller, on the contrary, the bigger induction range of width is bigger；The initialization width vector of RBF network isWherein,
Wherein, dji′For the i-th ' a value of j-th of width vector, i '=1,2 ..., n, n is the list of RBF neural input layer
First number, dfFor width adjusting coefficient, cji′Indicate the Center Parameter of radial basis function, N is given constant, and r is cyclic variable；
(2) center of radial basis function is determined
The Center Parameter of radial basis function is for completing the mapping of input layer to hidden layer；Each of hidden layer neuron is all
There is a radial basis function center, which is expressed as Cj=[Cj1, Cj2... Cjn]T, wherein
Wherein, q is the network concealed node layer number of RBF, miniWith maxiRespectively indicate all input letters of ith feature in training set
The minimum value and maximum value of breath；
(3) determine hidden layer to output layer weight
Weight W of the hidden layer to output layers=[ws1, ws2..wsp]T, s=1,2, wherein
Wherein, maxsWith minsRespectively indicate in training set all output informations i.e. average response time in s-th of output neuron
Maxima and minima.
7. the average response time appraisal procedure according to claim 6 towards seasonal form non-stationary concurrency, feature
It is: the output calculation formula of the network concealed layer of RBF described in step 3.2.2 are as follows:
The output calculation formula of output layer neuron are as follows:
O=[o1, o2]T (11)
Wherein, O is the output of output layer neuron.
8. the average response time appraisal procedure according to claim 7 towards seasonal form non-stationary concurrency, feature
It is: the step 3.3 method particularly includes:
Step 3.3.1: the history of the different moments of Web same on the virtual machine in cloud service system application respectively serviced is concurrent
Measure and inscribe when corresponding seasonal form in memory usage, CPU usage, average response time, and each service of prediction
Non-stationary concurrency data reach the input node of RBF network；
Step 3.3.2: using the center of the radial basis function of K-Means method initialization RBF network, stochastic gradient algorithm is used
The weight iteration of three network parameters in the RBF network of building is adaptively adjusted；
Step 3.3.3: three parameters of initialization RBF network: extension constant Dji, hidden layer radial basis function center DjiWith hidden
Connection weight W of the hiding layer to output layerkj；
Step 3.3.4: setting hidden layer interstitial content, Studying factors, model training error precision ε and maximum frequency of training MT；
Step 3.3.5: loop initialization the number of iterations cycle=1；
Step 3.3.6: based on the stochastic gradient descent method training big parameter of RBF network three, optimize the cloud service system based on RBF network
System average response time prediction model；
Step 3.3.7: the root-mean-square error RMSE of computation model training, if RMSE >=ε, i.e. RMSE is greater than a very small
Value, then cycle=cycle+1, no to then follow the steps 3.3.9；
Step 3.3.8: judging whether cycle is less than maximum frequency of training MT, if it is less, the 3.3.6 that gos to step, otherwise
Execute step 3.3.9；
Step 3.3.9: end loop obtains optimal average time response prediction model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910624505.8A CN110413657B (en) | 2019-07-11 | 2019-07-11 | Average response time evaluation method for seasonal non-stationary concurrency |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910624505.8A CN110413657B (en) | 2019-07-11 | 2019-07-11 | Average response time evaluation method for seasonal non-stationary concurrency |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110413657A true CN110413657A (en) | 2019-11-05 |
| CN110413657B CN110413657B (en) | 2021-08-17 |
Family
ID=68361083
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910624505.8A Active CN110413657B (en) | 2019-07-11 | 2019-07-11 | Average response time evaluation method for seasonal non-stationary concurrency |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110413657B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111367637A (en) * | 2020-03-03 | 2020-07-03 | 支付宝(杭州)信息技术有限公司 | Task processing method and device |
| CN111580955A (en) * | 2020-04-03 | 2020-08-25 | 上海非码网络科技有限公司 | Intelligent analysis system and method for computer utilization rate |
| WO2021237498A1 (en) * | 2020-05-27 | 2021-12-02 | Citrix Systems, Inc. | Load balancing of computing sessions with load patterns |
| CN116303094A (en) * | 2023-05-10 | 2023-06-23 | 江西财经大学 | Multipath coverage test method based on RBF neural network and individual migration |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106951471A (en) * | 2017-03-06 | 2017-07-14 | 浙江工业大学 | A kind of construction method of the label prediction of the development trend model based on SVM |
| CN108171379A (en) * | 2017-12-28 | 2018-06-15 | 无锡英臻科技有限公司 | A kind of electro-load forecast method |
| US20180307741A1 (en) * | 2017-04-25 | 2018-10-25 | Intel Corporation | Filtering training data for simpler rbf models |
-
2019
- 2019-07-11 CN CN201910624505.8A patent/CN110413657B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106951471A (en) * | 2017-03-06 | 2017-07-14 | 浙江工业大学 | A kind of construction method of the label prediction of the development trend model based on SVM |
| US20180307741A1 (en) * | 2017-04-25 | 2018-10-25 | Intel Corporation | Filtering training data for simpler rbf models |
| CN108171379A (en) * | 2017-12-28 | 2018-06-15 | 无锡英臻科技有限公司 | A kind of electro-load forecast method |
Non-Patent Citations (2)
| Title |
|---|
| 刘舒: "面向云服务性能自适应优化的组件副本动态增加方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 郭军等: "基于组件服务质量和服务性能的云服务性能瓶颈诊断方法", 《清华大学学报（自然科学版）》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111367637A (en) * | 2020-03-03 | 2020-07-03 | 支付宝(杭州)信息技术有限公司 | Task processing method and device |
| CN111367637B (en) * | 2020-03-03 | 2023-05-12 | 蚂蚁胜信(上海)信息技术有限公司 | Task processing method and device |
| CN111580955A (en) * | 2020-04-03 | 2020-08-25 | 上海非码网络科技有限公司 | Intelligent analysis system and method for computer utilization rate |
| WO2021237498A1 (en) * | 2020-05-27 | 2021-12-02 | Citrix Systems, Inc. | Load balancing of computing sessions with load patterns |
| US11586479B2 (en) | 2020-05-27 | 2023-02-21 | Citrix Systems, Inc. | Load balancing of computing sessions with load patterns |
| CN116303094A (en) * | 2023-05-10 | 2023-06-23 | 江西财经大学 | Multipath coverage test method based on RBF neural network and individual migration |
| CN116303094B (en) * | 2023-05-10 | 2023-07-21 | 江西财经大学 | Multipath coverage test method based on RBF neural network and individual migration |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110413657B (en) | 2021-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110413657A (en) | Average response time appraisal procedure towards seasonal form non-stationary concurrency | |
| CN108520357B (en) | Method and device for judging line loss abnormality reason and server | |
| CN104951425B (en) | A kind of cloud service performance self-adapting type of action system of selection based on deep learning | |
| CN106933650A (en) | load management method and system of cloud application system | |
| CN114616540A (en) | Autonomous cloud node range definition framework for big data machine learning use case | |
| US8090671B2 (en) | Dynamic performance modeling of computer application services using a serial parallel queueing network (SPQN) modeler | |
| CN108664378A (en) | A kind of most short optimization method for executing the time of micro services | |
| WO2008134143A1 (en) | Resource model training | |
| Menasce | Computing Missing Service Demand Parameters for Performance Models. | |
| CN109740052A (en) | The construction method and device of network behavior prediction model, network behavior prediction technique | |
| CN115422788B (en) | Power distribution network line loss analysis management method, device, storage medium and system | |
| Unuvar et al. | Selecting optimum cloud availability zones by learning user satisfaction levels | |
| CN114490065A (en) | Load prediction method, device and equipment | |
| CN109558248A (en) | A kind of method and system for the determining resource allocation parameters calculated towards ocean model | |
| CN108022123A (en) | The automatic adjusting method and device of a kind of business model | |
| CN112487612A (en) | Bus load combination prediction method and system based on distribution factor and information entropy | |
| Wang et al. | Adaptive placement of data analysis tasks for staging based in-situ processing | |
| EP4080789A1 (en) | Enhanced uncertainty management for optical communication systems | |
| CN112860531B (en) | Block chain wide consensus performance evaluation method based on deep heterogeneous graph neural network | |
| Naskos et al. | Elton: a cloud resource scaling-out manager for nosql databases | |
| Glavan et al. | Cloud environment assessment using clustering techniques on microservices dataset | |
| Tan et al. | Estimating parameters of a microsimulation model for breast cancer screening using the score function method | |
| Sinha et al. | Well-test model identification with self-organizing feature map | |
| Daraghmeh et al. | Incorporating Data Preparation and Clustering Techniques for Workload Segmentation in Large-Scale Cloud Data Centers | |
| CN116155835B (en) | Cloud resource service quality assessment method and system based on queuing theory |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |