CN104113590B - Copy selection method based on copy response time prediction - Google Patents

Abstract

The invention discloses a copy selection method based on copy response time prediction. Compared with an existing copy response time prediction algorithm, the method provided by the invention emphatically considers that the influence degree of sampling values of copy response time of different moments on prediction of copy response time is different. Closer is a sample to a prediction point, larger is a weight value, and otherwise, the weight value is smaller. Such a model solves the problem of different influence degree of influence factors on copy response time in different periods, thereby improving prediction precision.

Description

A kind of copy selection method predicted based on copy response time

Technical field

The present invention relates to the copy selection method in cloud environment, more particularly to a kind of pair predicted based on copy response time This system of selection.

Background technology

As the fast development of Internet technologies, internet data are sprayed into one's face such as tsunami, this directly reflects The relative plaque of mankind's processing data ability is weary.Currently, effective management method is needed in the face of abnormal huge data.This problem Core be that uncontrollable data increase and low resource utilization.

Cloud storage (cloud storage) be cloud computing (cloud computing) conceptive extension and develop The new ideas come, from once the focus for always research occur.Next Generation Internet is put forward higher requirement to storage, Storage system is encountered by the new challenge of the following aspects.(1) storage capacity requirement is bigger；(2) performance of storage system is higher； (3) storage system security level requirement is higher；(4) storage system is further intelligent.Cloud storage under such circumstances by Propose.In cloud storage system, data trnascription is its important ingredient.On the one hand the introducing of copy improves the peace of data Full property and availability, on the other hand, also bring the select permeability of optimum copy when copy is accessed.

The several factors for considering are needed in copy selection course：

1) distance：Copy requestor is apart from the distance between copy place host (hops)；

2) access speed of multihome node：Multihome node (host node) has availability (availability), can Stable provides service for copy requestor；Here availability include multihome node it is accessed when response speed, user When conducting interviews, first file to be read out from the memory node for have data trnascription, in the process for carrying out next step, So the access speed of memory node is very important.

3) bandwidth：The bandwidth of computer network refers to the maximum data rate that network can pass through, i.e., how much per secondBit.Data Bandwidth between copy place multihome node and user.Bandwidth is the primary factor that data trnascription is selected.Carrying out data trnascription During selection, the link for postponing minimum is usually selected to carry out data access.

4) calculating speed of calculate node：When copy selection is carried out, calculate node is entered by calculating inquiry locations of copies During row task scheduling, the disposal ability of calculate node is required to.

5) loading condition of multihome node：Identical data copy on a certain cloud storage node is accessed when there are multiple users When, the load of the node is also the another factor for affecting user's access speed.Duty factor may be carried as soon as possible when relatively low for user For service, when the node load for accessing is heavier, then the access request for needing user is ranked wait.

As can be seen here the selection influence factor of copy is numerous, to consider that process is also considerably complicated something one by one.On Five influence factors are stated, we can be reduced to one and consider a little in cloud environment：Response time (the Replica of data trnascription Response Time), i.e., send this period of time that solicited message starts to be met to end user's request from requestor. Therefore, if the copy response time for predicting following the N+1 time can be passed through, N represents that copy has had the access note of history n times Record, user is just referred to the response time of copy, selects copy to respond most short copy and conducts interviews.

At present, researcher also has the example that the response time to copy is predicted, and it is sent out by way of replica location All optional data trnascription needed for existing user.User node dynamic arranges the history of the nearest n times of optional data trnascription and accesses Information, the copy response time of the N+1 time is predicted by setting up simplest grey forecasting model.But such mode is simultaneously Do not take into full account that copy response time each time, for the difference of the influence degree of prediction, entirely predicts imprecision.Because The closer to future position copy response time sample value for the influence degree of the prediction of user behavior be bigger, it is impossible to without exception In.

The content of the invention

To solve above-mentioned technical problem, the present invention is adopted the following technical scheme that：

A kind of copy selection method predicted based on copy response time, it is comprised the steps of：

Step one, for the same data trnascription that user's request is crossed, judge whether there is the data pair of n*N time in server This access record, n is the total time hop count at the place for sampling sample, and N is that the acquirement n times copy in each sampling time period rings Answer time value.If it has, going to step two；Otherwise go to step five；

Step 2, n different time sections for taking main server-side, each time period takes n times copy response time, calculates Each time to the response time of same copy, { t is obtained_{i.j.R_1}、t_{i.j.R_2}、…、t_{i.j.R_N}, t_{i.j.R_N}Represent the number to file i According to the copy response time that the n-th of the copy R of block j is accessed；

Step 3, according to formula (1) calculate y_t, i.e. the copy response time of future position t the N+1 time；

y_t=β₀(t)+β₁(t)x₁+β₂(t)x₂+,...+β_f(t)x_f (1)

F represents that copy is accessed for the f time, f ∈ { 1,2,3..., N }, and N represents that each sampling time period takes altogether n times copy sound Answer time value；β_fCopy response time when () represents that t copy is accessed for the f time t is to the N+1 time copy of future position The impact weights of response time, under different f, the value of β is different；x_fRepresent copy SF_i.DB_j.Replica_rIn the f time quilt Copy response time during access；

From future position t more close to copy response time sampling value it is bigger to predicted impact, otherwise, affect it is relatively small A bit.W (t) wherein in formula (2) adjusting weighing factor, to weigh impact journey of the sampling value of different time sections to future position Degree.Weighted least square is exactly to seek β₀,β₁,...,β_fSo that formula (2) sum of deviation square Q_tIt is minimum.

x_nfRepresent copy response time value of the copy copy when the f time copy of n-th sampling time period is accessed. β_fImpact power of copy response time when representing that copy is accessed for the f time to the response time of the N+1 time copy of future position Value.

Step 4, above step one～tri-, for some source data SF_i.DB_jSome data trnascription SF_i.DB_j.Replica_R, the history n time period is sampled, each time period takes the copy response time value of n times, asks pre- The N+1 time copy response time t of measuring point t_{i.j.R_N+1}.Repeat above step one～tri-, calculate same source data not With the copy response time t of copy_{i.j.1_N+1}, t_{i.j.2_N+1}..., t_{i.j.R_N+1}..., t_{i.j.r_N+1}, R ∈ { 1,2 ..., r }, r are SF_i.DB_jCopy sum.Select t_{i.j.R_N+1}Minimum copy conducts interviews；

If creating a Copy and conducting interviews without this copy in step 5, master server；Otherwise select accessed most frequency Numerous copy conducts interviews, i.e., most frequent copy occur in the timestamp record of master server.

Copy response time computational methods in step 2 are as follows：

Step 1, request message message that data block j for file i is sent with user orientation server_{i.j_x(request)}, (x It is user to data block SF_i.DB_jX: th is accessed), the timestamp of message is generated with regard to this：Timestamp_{i.j_x(reqmessage)}；

Step 2, server are connected to subscription request message, and the timestamp of request message is recorded.Analysis request message And the copy of corresponding source data is inquired, inquiry just follows original route and replies message to user after finishing message_{i.j_x(answer)}, with the details of related copy in message；

Step 3, user receive the feedback that server is beamed back, and just select the multihome node on server to conduct interviews, and visit Ask after finishing, to server confirmation message m essage is sent_{i.j.R_x(received)}.Now server is again by the corresponding time Stamp records Timestamp_{i.j.R_x(receivedmessage)}。

Step 4, server end by each copy it is corresponding each time be accessed when Timestamp_{i.j_x(reqmessage)}With Timestamp_{i.j.R_x(receivedmessage)}Its difference is sought, the response time t of copy is just obtained_{i.j.R_x}：

t_{i.j.R_x}=Timestamp_{i.j.R_x(receivedmessage)}-Timestamp_{i.j_x(reqmessage)}。 (3)

Step 5, based on the above the step of, the copy response time of the different copy history n times of same data block just by Record in server end.t_{i.j.r_x}Represent that x: th accesses SF_i.DB_j.Replica_rResponse time.

The concrete solution procedure of above-mentioned steps three is as follows：

Order

Then X β=Y,Y is the N+1 time copy response time of each time period of n time period of sampling Actual value column matrix.Represent the estimated value f ∈ { 1,2,3..., N }, X of t the N+1 time copy response time of copy T () is observed values of the X in t, copy response time is to future position t when β (t) represents that the copy of t n times is accessed The impact weights of the response time of the N+1 time copy.

Q_tThe sum of deviation square of representation formula (2), x_nfRepresent the f time copy of the copy in the sample point of n-th time period The response time value of copy when accessed.β_fThe N+1 time to future position of copy response time when representing that the f time copy is accessed The impact weights of copy response time.

By

δ β (t) represents the deviation of t copy sampling value,Expression seeks first derivative to δ β (t).

Equation can be obtained

X^TW (t) X β (t)=X^TW(t)Y (7)

Equation has and does not uniquely solve, and its arbitrary solution is exactlyRepresent that the f copy response time that t is taken affects power Value matrix.Obtain

Represent the average deviation quadratic sum of n sampling time period.

The computational methods of different sampling time period weighing factors are as follows in step 3：

W (t)=diag (w₁(t),w₂(t),...,w_n(t)), W (t) represent taken n copy sampling time period each Weighing factor matrix of the sampling value of time period to future position copy response time, detailed process is as follows：

Order

Ask for here with following methods methodOrder

1) assume that θ is quantitative

2) calculated by formula (8)

Wherein X (quf)^T, Y (quf) is X, Y remove the data after f row data.Can be obtained according to formula (1)X (t) is the t row vectors of X, The error of actual value and predictive value we be expressed as

3) calculateRepresent the difference of actual value and predictive value；

4) calculate θ so that CV (θ)=min (CV), that is, in the case of causing actual value minimum with the error of prediction, ask for θ values now.

The prediction of response time is more when the copy response time prediction algorithm of the present invention causes following copy accessed Science, compared to existing copy response time prediction mode, the present invention considers emphatically the pair of range prediction point different time sections This response time sampling value is different to future position influence degree.Use weighting mode cause from future position more close to sample shadow Ring weighted value bigger, therefore solve that influence factor is different to the influence degree of copy response time in the period of different to be asked Topic, improves the precision of prediction.

Meanwhile, propose a kind of time difference by server end record subscription request message and confirmation message time stamp to count Calculate the response time of history copy.

Description of the drawings

The interaction diagrams of Fig. 1 server ends and client

Fig. 2 copy response time prognostic charts

The contrast of the response time of copy one under two kinds of prediction modes of Fig. 3

The contrast of the response time of copy two under two kinds of prediction modes of Fig. 4

The contrast of the response time of copy three under two kinds of prediction modes of Fig. 5

Specific embodiment

The technical problem to be solved is how the copy response time of the analysis of history n times of science is so as to pre- Survey the copy response time of the N+1 time.There is provided a kind of copy selection algorithm predicted based on copy response time, it is contemplated that every time Access difference of the copy response time to a following copy response time predicted impact degree, the statistical data to history n times It is weighted, and will predicts the outcome and feed back to user, the copy that user selects copy response time most short conducts interviews

Technical scheme is described in detail below in conjunction with the accompanying drawings：

As shown in figure 1, setting SF_i.DB_j.Replica_rRepresent the copy r of data block j of file i, server end and client Using the message transfer mechanism of active facilitate master server (Master Server, MS) to copy response time statistics.

If master server has the details that special copy response time statistical module is asked to counting user.System In constantly have user to conduct interviews copy, as { 1,2 ..., x } is secondary, user's request person's x: th to server send it is right In request message message of data block j of file i_{i.j_x(request)}, server receive request after record first user send please Seek the timestamp attribute Timestamp at message moment_{i.j_x(reqmessage)}, x is user to data block SF_i.DB_jX: th is accessed, it User's request being responded afterwards and analyzing and inquire about qualified multihome node, to user response message is beamed back message_{i.j.r_x(answer)}, in message comprising meet condition multihome node specifying information, user receive after on server Multihome node conduct interviews, request be met after with user orientation server beam back confirmation message m essage_{i.j.r_x(received)}, Server receives confirmation message m essage_{i.j.r_x(received)}The timestamp of registration confirmed message at once afterwards Timestamp_{i.j.r_x(receivedmessage)}。

The interaction that here is come between analog subscriber server using the agreement similar to three-way handshake in network：

Step 1, request message message that data block j for file i is sent with user orientation server_{i.j_x(request)}, x Represent user to data block SF_i.DB_jX: th is accessed, and the timestamp of message is generated with regard to this：Timestamp_{i.j_x(reqmessage)}

Step 2, server are connected to subscription request message, and the timestamp of request message is recorded.Analysis request message And the copy of corresponding source data is inquired, inquiry just follows original route and replies message to user after finishing message_{i.j_x(answer)}, details of the message with related copy；

Step 3, user receive the feedback that server is provided, and just select the multihome node on server to conduct interviews, and visit Ask after finishing, to server confirmation message m essage is beamed back_{i.j.R_x(received)}.Now server is again by the corresponding time Stamp is recorded：Timestamp_{i.j.R_x(receivedmessage)}。

Step 4, server end by each copy it is corresponding each time be accessed when Timestamp_{i.j_x(reqmessage)}With Timestamp_{i.j.R_x(receivedmessage)}Its difference is sought, the response time t of copy is just obtained_{i.j.R_x}：

t_{i.j.R_x}=Timestamp_{i.j.R_x(receivedmessage)}-Timestamp_{i.j_x(reqmessage)}。

Step 5, based on the above the step of, the copy response time of the different copy history n times of same data block just by Record in server end.t_{i.j.r_x}Represent that x: th accesses SF_i.DB_j.Replica_rResponse time.

1 is shown in Table in detail：

The copy access information of table 1

Copy id	The response time t of copyi.j.r.frquency
		SFi.DBj.Replica1	ti.j.1_1
SFi.DBj.Replica1	ti.j.1_2
		…	…
SFi.DBj.Replica1	ti.j.1_x
		SFi.DBj.Replica2	ti.j.2_1
SFi.DBj.Replica2	ti.j.2_2
		…	…
SFi.DBj.Replica2	ti.j.2_x
		…	…
SFi.DBj.Replicar	ti.j.r_1
		SFi.DBj.Replicar	ti.j.r_2
SFi.DBj.Replicar	…
		SFi.DBj.Replicar	ti.j.r_x

Work as SF_i.DB_jIt is requested by a user for the N+1 time, the copy response time { t of master server statistical history n times_{i.j.R_1}、 t_{i.j.R_2}、…t_{i.j.R_N}、…、t_{i.j.r_N}, R ∈ { 1,2 ..., r } herein, r is SF_i.DB_jCopy sum, using prediction algorithm The response time t that prediction copy is the N+1 time_{i.j.R_N+1}.When same source data has different copies, statistics is all Response time { the t of copy_{i.j.1_N+1},t_{i.j.2_N+1},...t_{i.j.R_N+1},...,t_{i.j.r_N+1}}.User is selected in these copies t_{i.j.R_N+1}The minimum copy of value conducts interviews.

As shown in Fig. 2 pre- flow gauge is as follows：

Step one, for the same data trnascription that user's request is crossed, judge whether there is the data pair of n*N time in server This access record, n is the total time hop count at the place for sampling sample, and N is that the acquirement n times copy in each sampling time period rings Answer time value.If it has, going to step two；Otherwise go to step five；

Step 2, n different time sections for taking main server-side, each time period takes n times copy response time, calculates Each time to the response time of same copy, { t is obtained_{i.j.R_1}、t_{i.j.R_2}、…、t_{i.j.R_N}, t_{i.j.R_N}Represent the number to file i According to the copy response time that the n-th of the copy R of block j is accessed；

Step 3, according to formula (1) calculate y_t, i.e. the copy response time of future position t the N+1 time；

y_t=β₀(t)+β₁(t)x₁+β₂(t)x₂+,...+β_f(t)x_f (1)

F represents that copy is accessed for the f time, f ∈ { 1,2,3..., N }, and N represents that each sampling time period takes altogether n times copy sound Answer time value；β_fCopy response time when () represents that t copy is accessed for the f time t is to the N+1 time copy of future position The impact weights of response time, under different f, the value of β is different；x_fRepresent copy SF_i.DB_j.Replica_rIn the f time quilt Copy response time during access；

From future position t more close to copy response time sampling value it is bigger to predicted impact, otherwise, affect it is relatively small A bit.W (t) wherein in formula (2) adjusting weighing factor, to weigh impact journey of the sampling value of different time sections to future position Degree.Weighted least square is exactly to seek β₀,β₁,...,β_fSo that formula (2) sum of deviation square Q_tIt is minimum.

x_nfRepresent copy response time of the copy when the f time copy of the sample point of n-th sampling time period is accessed Value.β_fRepresent the f time it is accessed when impact of the copy response time to the response time of the N+1 time copy of future position weigh Value.

The following is the concrete solution procedure of step 3：

Order

Then X β=Y,Y is the N+1 time copy response time of each time period of n time period of sampling Actual value matrix.Represent the estimated value (f ∈ { 1,2,3..., N }) of t the N+1 time copy response time of copy, X (t) For X t observed value, copy response time is to future position t N when β (t) represents that the copy of t n times is accessed The impact weights of the response time of+1 copy.

Q_tThe sum of deviation square of representation formula (2), x_nfRepresent the f time copy of the copy in the sample point of n-th time period The response time value of copy when accessed.β_fThe N+1 time to future position of copy response time when representing that the f time copy is accessed The impact weights of copy response time.

By

δ β (t) represents the deviation of t copy sampling value,Expression seeks first derivative to δ β (t).

Equation can be obtained

X^TW (t) X β (t)=X^TW(t)Y (6)

Equation has and does not uniquely solve, and its arbitrary solution is exactlyRepresent the f copy response time weights that t is taken Matrix.

Obtain

Represent the average deviation quadratic sum of n sampling time period.

The calculating of different sampling time period weighing factors is presented herein below：

W (t)=diag (w₁(t),w₂(t),...,w_n(t)), W (t) represent taken n copy sampling time period each Weighing factor matrix of the sampling value of time period to future position copy response time, detailed process is as follows：

Order

Ask for here with following methods methodOrder

1) assume that θ is quantitative

2) calculated by formula (7)

Wherein X (quf)^T, Y (quf) is X, Y remove the data after f row data.Can be obtained according to formula (1)X (t) is the t row vectors of X, The error of actual value and predictive value we be expressed as

3) calculateRepresent the difference of actual value and predictive value；

4) θ is calculated so that CV (θ)=min (CV) so that ask for now in the case that actual value is minimum with the error of prediction θ values.

Step 4, above step one～tri-, for some source data SF_i.DB_jSome data trnascription SF_i.DB_j.Replica_R, the history n time period is sampled, each time period takes the copy response time value of n times, asks pre- The N+1 time copy response time t of measuring point t_{i.j.R_N+1}.Repeat above step one～tri-, calculate same source data not With the copy response time t of copy_{i.j.1_N+1}, t_{i.j.2_N+1}..., t_{i.j.R_N+1}..., t_{i.j.r_N+1}, R ∈ { 1,2 ..., r }, r are SF_i.DB_jCopy sum).Select t_{i.j.R_N+1}Minimum copy conducts interviews.

If without this copy in step 5, master server, first creating a Copy and then conducting interviews；Otherwise select accessed Most frequent copy conducts interviews, i.e., most frequent copy occur in the timestamp record of master server.

The effectiveness of the method is verified below by way of specific embodiment：

1st, assume j-th data block of i-th file, 3 copies of this data have been coexisted on server node.

2nd, future position is moment t, takes range prediction point 1 time period, 2 time periods, 3 time periods：T-T, t-2T, t- 3T is used as data point.And take the nearest 4 time copy access information at this three moment respectively to predict the pair of the 5th This response time, (the 0th time is shown as in analogous diagram to the 4th), according to copy request time recorded request time is stabbed To calculate the response time of copy.

We use two ways：

It is a kind of be the modes for considering that not sample value in the same time is different to the impact of predictive value i.e. set forth herein scheme；

Another kind is not consider not influence degree of the sample value predictive value to copy in the same time, and is used uniformly across identical sample This impact weights；

We are given to predicting the outcome using Matlab emulation tools and are intuitively described, and are predicted the outcome and are seen Fig. 3, Fig. 4, Fig. 5；

Experimental analysiss：The prediction of the copy response time of 3 copies of same data source more than, it can be seen that make With set forth herein copy response time prediction mode prediction the more close copy of copy response time t it is actual Copy response time.Predicting the outcome for copy response time has directly influenced the selection of copy.

From lower Fig. 3,4,5 it can be seen that because the impact of precision of prediction, when t copy the 5th is accessed, such as according to The predicting the outcome of common average weight can cause user to select copy 1 or 2, and use according to the close future position of sample value away from Just can accurately judge that copy 2 is the copy to be selected from the mode for taking impact weights.

Can illustrate from this point, when copy response time prediction algorithm proposed by the present invention causes following copy to be accessed The prediction of response time more science, compared to existing copy response time prediction mode, the present invention considers more to connect emphatically The copy response time sample influence degree of nearly time at present is higher, and precision of prediction is more increased.

Claims (4)

1. it is a kind of based on copy response time predict copy selection method, it is comprised the steps of：

Step one, for the same data trnascription that user's request is crossed, the data trnascription for judging whether to have n*N time in server Record is accessed, n is the total time hop count at the place for sampling sample, when N is that the acquirement n times copy in each sampling time period is responded Between be worth；If it has, going to step two；Otherwise go to step five；

Step 2, n different time sections for taking main server-side, each time period takes n times copy response time, calculates each The secondary response time to same copy, obtains { t_{i.j.R_1}、t_{i.j.R_2}、…、t_{i.j.R_N}, t_{i.j.R_N}Represent the data block to file i The copy response time that the n-th of the copy R of j is accessed；

Step 3, according to formula (1) calculate y_t, i.e. the copy response time of future position t the N+1 time；

y_t=β₀(t)+β₁(t)x₁+β₂(t)x₂+,...+β_f(t)x_f (1)

F represents that copy is accessed for the f time, f ∈ { 1,2,3..., N }, when N represents that each sampling time period takes altogether the response of n times copy Between be worth；β_fResponse of copy response time when () represents that t copy is accessed for the f time t to the N+1 time copy of future position The impact weights of time, under different f, the value of β is different；x_fRepresent copy SF_i.DB_j.Replica_rIt is accessed at the f time When copy response time；

From future position t more close to copy response time sampling value it is bigger to predicted impact, otherwise, affect it is relatively small；Its W (t) in middle formula (2) adjusting weighing factor, to weigh the influence degree of the sampling value to future position of different time sections；Plus Power least-squares estimation is exactly to seek β₀,β₁,...,β_fSo that formula (2) sum of deviation square Q_tIt is minimum；

$Q_t\left({\mathrm{\β}}_0\left(t\right),{\mathrm{\β}}_1\left(t\right),\mathrm{...},{\mathrm{\β}}_f\left(t\right)\right)=\underset{t=1}{\overset{n}{\mathrm{\Σ}}}{w}_t{\left(y_t-{\mathrm{\β}}_0\left(t\right)-{\mathrm{\β}}_1\left(t\right)x_{11}-\mathrm{...}-{\mathrm{\β}}_f\left(t\right)x_{nf}\right)}^2---\left(2\right)$

x_nfRepresent copy response time value of the copy when n-th sampling time period, the f time copy is accessed；β_fRepresent copy Impact weights of copy response time when being accessed for the f time to the response time of the N+1 time copy of future position；

Step 4, above step one～tri-, for some source data SF_i.DB_jSome data trnascription SF_i.DB_j.Replica_R, the history n time period is sampled, each time period takes the copy response time value of n times, asks pre- The N+1 time copy response time t of measuring point t_{i.j.R_N+1}；Repeat above step one～tri-, calculate same source data not With the copy response time t of copy_{i.j.1_N+1}, t_{i.j.2_N+1}..., t_{i.j.R_N+1}..., t_{i.j.r_N+1}, R ∈ { 1,2 ..., r }, r are SF_i.DB_jCopy sum；Select t_{i.j.R_N+1}Minimum copy conducts interviews；

If creating a Copy and conducting interviews without this copy in step 5, master server；Otherwise select accessed most frequent Copy conducts interviews, i.e., most frequent copy occur in the timestamp record of master server.

2. it is according to claim 1 it is a kind of based on copy response time predict copy selection method, in its step 2 Copy response time computational methods are as follows：

Step 1, request message message that data block j for file i is sent with user orientation server_{i.j_x(request)}, x represents User is to data block SF_i.DB_jX: th is accessed, and the timestamp of message is generated with regard to this：Timestamp_{i.j_x(reqmessage)}；

Step 2, server are connected to subscription request message, and the timestamp of request message is recorded；Analysis request message is simultaneously looked into The copy of corresponding source data is ask, inquiry just follows original route and replies message message to user after finishing_{i.j_x(answer)}, disappear With the details of related copy in breath；

Step 3, user receive the feedback that server is beamed back, and just select the multihome node on server to conduct interviews, and have accessed Bi Hou, to server confirmation message m essage is sent_{i.j.R_x(received)}；Now server is again by corresponding time stamp Record Timestamp_{i.j.R_x(receivedmessage)}；

Step 4, server end by each copy it is corresponding each time be accessed when Timestamp_{i.j_x(reqmessage)}With Timestamp_{i.j.R_x(receivedmessage)}Its difference is sought, the response time t of copy is just obtained_{i.j.R_x}：

t_{i.j.R_x}=Timestamp_{i.j.R_x(receivedmessage)}-Timestamp_{i.j_x(reqmessage)}； (3)

Step 5, based on the above the step of, the copy response time of the different copy history n times of same data block is just recorded Get off in server end；t_{i.j.r_x}Represent that x: th accesses SF_i.DB_j.Replica_rResponse time.

3. it is according to claim 1 it is a kind of based on copy response time predict copy selection method, the tool of its step 3 Body solution procedure is as follows：

Order

Then X β=Y,Y is the reality of the N+1 time copy response time of each time period of n time period of sampling The column matrix of actual value；Represent that the estimated value f ∈ { 1,2,3..., N } of t the N+1 time copy response time of copy, X (t) are , in the observed value of t, copy response time is to future position t N+1 when β (t) represents that the copy of t n times is accessed for X The impact weights of the response time of secondary copy；

$Q_t=\underset{t=1}{\overset{n}{\Σ}}{w}_t{(y_t-{\mathrm{\β}}_0(t)-{\mathrm{\β}}_1(t)x_{11}-...-{\mathrm{\β}}_f(t\left)x_{nf}\right)}^2={(Y-X\mathrm{\β}(t\left)\right)}^T*W\left(t\right)(Y-X\mathrm{\β}(t\left)\right)---\left(5\right)$

Q_tThe sum of deviation square of representation formula (2), x_nfWhen representing copy the f time copy being accessed in n-th sampling time period Copy response time value；β_fCopy response time is responded to the N+1 time copy of future position when representing that the f time copy is accessed The impact weights of time；

By

$\∂\frac{Q\[\mathrm{\β}\left(t\right)\]}{\mathrm{\δ}\mathrm{\β}\left(t\right)}=-2X^{T}W\left(t\right)Y+2X^{T}W\left(t\right)X\mathrm{\β}\left(t\right)=0---\left(6\right)$

δ β (t) represents the deviation of t copy sampling value,Expression seeks first derivative to δ β (t)；

Equation can be obtained

X^TW (t) X β (t)=X^TW(t)Y (7)

Equation has and does not uniquely solve, and its arbitrary solution is exactlyRepresent that the f copy response time that t is taken affects weights square Battle array；

Obtain

Represent the average deviation quadratic sum of n sampling time period.

4. difference takes in a kind of copy selection method step 3 based on the prediction of copy response time according to claim 3 The sample time period computational methods of weighing factor are as follows：

W (t)=diag (w₁(t),w₂(t),...,w_n(t)), W (t) represents n taken each time of copy sampling time period Weighing factor matrix of the sampling value of section to future position copy response time, detailed process is as follows：

Order

Ask for here with following methods methodOrder

1) assume that θ is quantitative

2) calculated by formula (8)

Wherein X (quf)^T, Y (quf) is X, Y Remove the data after f row data；Can be obtained according to formula (1)X (t) is the t row vectors of X, actual The error of value and predictive value we be expressed as

3) calculateRepresent the difference of actual value and predictive value；

4) calculate θ so that CV (θ)=min (CV), that is, in the case of causing actual value minimum with the error of prediction, ask for now θ values.