CN107241442B - A kind of key assignments data storage storehouse copy selection method based on prediction - Google Patents
A kind of key assignments data storage storehouse copy selection method based on prediction Download PDFInfo
- Publication number
- CN107241442B CN107241442B CN201710631076.8A CN201710631076A CN107241442B CN 107241442 B CN107241442 B CN 107241442B CN 201710631076 A CN201710631076 A CN 201710631076A CN 107241442 B CN107241442 B CN 107241442B
- Authority
- CN
- China
- Prior art keywords
- waiting list
- list length
- feedback information
- copy
- length
- 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.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
- H04L67/1001—Protocols in which an application is distributed across nodes in the network for accessing one among a plurality of replicated servers
- H04L67/1004—Server selection for load balancing
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/01—Protocols
- H04L67/10—Protocols in which an application is distributed across nodes in the network
- H04L67/1097—Protocols in which an application is distributed across nodes in the network for distributed storage of data in networks, e.g. transport arrangements for network file system [NFS], storage area networks [SAN] or network attached storage [NAS]
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/60—Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources
- H04L67/63—Routing a service request depending on the request content or context
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
This patent discloses a kind of key assignments data storage storehouse copy selection method based on prediction.Including waiting list length change Trend judgement, prediction waiting list length and copy three steps of sequence performed successively.When being predicted, PRS is judged the variation tendency for waiting queue length first, then screens feedback information according to the variation tendency of waiting list length, to avoid out-of-date feedback information from influenceing prediction result；Secondly, RPS chooses the waiting list length estimate that corresponding method carries out replica server according to the trend of current waiting list length change；Finally, marking sequence is carried out to replica server according to the waiting list length of estimation, selects most suitable replica server.By prediction, PRS can be obtained than C3 more accurately replica server waiting list length estimations, so as to make more preferable copy trade-off decision, further reduce the response time of application request.
Description
Technical field
The present invention relates to a kind of key assignments data storage storehouse copy selection method based on prediction.
Background technology
In distributed key assignments storage system, a terminal request can generate substantial amounts of key assignments, it is necessary to tens even
Hundreds of servers are serviced.The delay of these key assignments request directly affects the response time of terminal request.Response time
One of long an important factor for always influenceing Consumer's Experience and income, the tail delay for reducing key assignments request are significant.
Wherein, it is one of effective ways of reduction tail delay for the suitable replica server of key assignments request selecting.
The reasons such as the performance inconsistency due to server, choose suitable replica server not a duck soup.First, it is if all
Client all select most fast replica server, high concurrent can cause the drastically decline of replica server performance.Therefore, copy
Selection algorithm allows for avoiding such Herd Behavior.Secondly as the light weight level characteristics of key assignments access, copy selection algorithm
Must be simple enough.To solve problem above, the performance change of adaptation server, copy selection algorithm C3 arises at the historic moment.C3's
Main thought is that with return value are fed back into client the service time of the waiting list length information of server and key assignments
End, and marking sequence is carried out to replica server by these feedback informations, so as to select most fast replica server.
In key assignments data storage storehouse, copy selection algorithm determines the replica server of each key assignments accessing operation, pole
The delay of each key assignments accessing operation of earth effect, it is that response time for reducing application request, (that is, all key assignments operated
Tail is delayed) one of important method.Although C3 is showed well in terms of tail delay is reduced, but it have been found that C3 still exists
The room for improvement of one step.When receiving feedback information, C3 carries out exponential weighting smooth (EWMA) to feedback information and calculated to estimate
The waiting list length of server end and the service rate of server.However, copy sorting operation and receiving between feedback information
Existence time is poor, once this stand-by period (τw) longer, the performance inconsistency of server can cause the state of server to become
Change, and waiting list length estimation before will be unable to react server state now.We are referred to as such case
Timeliness sex chromosome mosaicism.Although introducing feedback information creative current copy selection algorithm C3 instructs copy to select, and obtains
Good performance, but it, which has the ageing of feedback information, often becomes very poor.As ageing very poor (τw≥
When 100ms), waiting list length estimate inaccurate C3, the selection of replica server will be directly affected, cause waiting list to grow
Rapid development is spent, and then causes queueing delay to increase severely.
The content of the invention
In order to solve the timeliness sex chromosome mosaicism of above-mentioned C3 copies selection algorithm, the present invention proposes a kind of key assignments based on prediction
Data storage storehouse copy selection method, i.e. Predication-based Replica SelectionAlgorithm, referred to as
PRS.When feedback information does not have ageing, i.e. τwDuring >=100ms, PRS methods replace simple index using Forecasting Methodology
The waiting list length of the smooth estimation server of weighting.It is more suitably secondary by more accurately waiting list length estimate, selection
Book server.
In order to realize above-mentioned technical purpose, the technical scheme is that,
A kind of key assignments data storage storehouse copy selection method based on prediction, including the waiting list length performed successively become
Change Trend judgement, prediction waiting list length and copy three steps of sequence；
Wherein waiting list length Trend judgement includes following three steps：
Step 1：Will be including waiting list length change amount by way of the return value carrying that key assignments accesses operation
Feedback of the information to client, when client receives feedback information, perform step 2；
Step 2：The variation tendency of the waiting list length of current copy server is judged according to feedback information；
Step 3：If waiting list length change trend does not change, current feedback information and its timestamp are stored, if
Waiting list length change trend changes, and stores current feedback information and its timestamp, while deletes a wait team
The feedback information of all storages in row length change trend；
Predictive server waiting list length is divided into following three steps：
Step 1：The ageing power of the feedback information newly stored in judgment step three, if ageing by force perform step
Rapid 2, if weak ageing execution step 3；
Step 2：The waiting list length in this strong ageing feedback information is taken, with the feedback letters received all before
Waiting queue length to carry out the smooth historical record result of exponential weighting in breath, to carry out exponential weighting again smooth, to be used as server
The waiting list length estimate at end, and result is stored as the smooth historical record result of new exponential weighting, then perform pair
This sorting operation；
Step 3：The information stored according to the waiting list length change trend and step 3 of step 2, to server end
Waiting list length be predicted, perform copy sorting operation；
Copy sorting operation：The estimation waiting list length calculated according to step 2 or step 3 is to server copy progressive
It can assess, finally choose the best replica server of performance evaluation result.
Described method, in step 1, feedback information includes the variation delta q of the waiting list length in per 5mss, etc.
Treat queue lengthAnd the service time μ that current key assignments accessess。
Described method, in step 2, according to the variation delta q of waiting list lengthsWith waiting list lengthWill etc.
Treat that the variation tendency of queue length is divided into three kinds of states, steady, raising and lowering；Original state is arranged to plateau, three kinds
The Rule of judgment changed between state is as follows：In any case ifThen it is judged as plateau；From steady shape
State, which is converted to propradation, need to meet Δ qs＞ 0 withSimultaneouslyChanged from plateau
It need to meet Δ q for decline states＜ 0 withSimultaneouslyDecline is converted to from propradation
State need to meet Δ qs＜ 0 orSimultaneouslyNeeded from State Transferring is declined for propradation
Meet Δ qs＞ 0 orSimultaneously
Described method, in step 3, if waiting list length change trend changes, by following steps come
Judge whether a upper feedback information retains：New feedback information is subtracted using the waiting list length in a upper feedback information
Middle wait queue length, if result is just, and current waiting list length change trend is declines, the feedback information needs to protect
Stay；If result is negative, and current waiting list length change trend is rises, the feedback information is also required to retain, otherwise not
Retain.
Described method, in step 1, judge the whether ageing power of feedback information be with the timestamp of feedback information with
Current time compares, and threshold value is set as 100ms, if less than equal to 100ms, feedback information have it is strong ageing, if greater than
100ms, feedback information have weak ageing.
Described method, in step 3, if waiting list length change trend is rises or falls state, prediction is used
Method be least-square fitting approach, according to currently stored feedback waiting list length and its timestamp, fit one
Waiting list length substitutes into current time, result of calculation is the waiting list length predicted to the linear function of time；
If waiting list length change trend is plateau, exponential weighting smoothing algorithm prediction waiting list length is usedCalculation formula is as follows：
WhereinBy step 3 in the same waiting list length change trend stored at present, put down using exponential weighting
Sliding algorithm is predicted the historical record of waiting list length.
Described method, in copy sorting operation, following steps are divided into the operation that copy is ranked up：By step 2 or 3
The waiting list length value calculated substitutes into scoring functions, and copy is ranked up according to the order of score from small to large, chooses and divides
The minimum replica server of number.Scoring functions are as follows：
Wherein, RsThe response time collected for client,For the EWMA values of the service time of feedback information, n is client
Number is held, OSK is the client corresponding server s key assignments access request number for having sent but not responded.
The technical scheme that the present invention solves above-mentioned timeliness sex chromosome mosaicism includes following steps：First, waiting list is grown
The variation tendency of degree is divided into three kinds of states：Plateau, propradation and decline state.When receiving feedback information, by anti-
Feedforward information judges the trend of the current waiting list length change of server, and by the feedback information in current trend and at that time
Between stab and store, by the feedback letter of upper waiting list length change trend in addition to the useful feedback information positioned at border
Breath is deleted.Secondly, feedback information is divided into τ by ageingwMore than 100ms and no more than two parts of 100ms.Experiment proves τw
Feedback information no more than 100ms is with ageing available feedback information, therefore, in τwDuring no more than 100ms, still adopt
With with C3 identical waiting list length estimate methods.In τwDuring more than 100ms, pass through the change of current waiting list length
Trend and the feedback information stored after waiting list length change Trend judgement are carried out to the waiting list length of server
Prediction.After being estimated to server end waiting list length, PRS still carries out copy from C3 identicals scoring functions
Sequence, but because PRS waiting list length prediction is more accurate, therefore PRS performance is better than C3.
The invention will be further described below in conjunction with the accompanying drawings.
Brief description of the drawings
Fig. 1 is that the copy of key assignments storage system selects Organization Chart.
Fig. 2 is C3 waiting list length, feeds back waiting list length and estimation waiting list length, upper figure is τw≥
100ms situation, figure below τw＜ 100ms situation.
Fig. 3 is C3 waiting list length and its waiting list length of feedback.
Fig. 4 is PRS waiting list length trend and its Rule of judgment schematic diagram.
For Fig. 5 when waiting list length change trend changes, the feedback information positioned at border is not required to situation about deleting
Schematic diagram.
Fig. 6 is PRS prediction waiting list length and the design sketch of waiting list length Trend judgement.
Fig. 7 is the test design sketch of delay performance of the PRS algorithms with C3 algorithms under different clients number environment.
Fig. 8 is test design sketch of the PRS algorithms with C3 algorithms in the environment of 20% client produces 80% request.
Fig. 9 is test design sketch of the PRS algorithms with C3 algorithms in the environment of 50% client produces 80% request.
Figure 10 is the test design sketch of delay performance of the PRS algorithms from C3 algorithms in the environment of different number of copies.
Embodiment
To make the purpose of the present invention, content and advantage clearer, embodiments of the present invention are entered below in conjunction with the accompanying drawings
Row is specific to be illustrated.
PRS algorithms of the present invention, it is the timeliness sex chromosome mosaicism for copy selection algorithm C3, is improved.Due to row
Team's delay is the important component of tail delay, so waiting list length estimate is the core place of copy scoring functions.C3
Copy selection algorithm is in τwDuring more than 100 milliseconds, serious timeliness sex chromosome mosaicism can be run into.Now, using out-of-date feedback information
It is subject to the estimation waiting list length that exponential weighting smooth (EWMA) obtains and differs very big with the waiting list length of server.Such as
Shown in Fig. 2, QsThe true waiting list length of server end is represented,Represent the feedback waiting list length after EWMA.
As can be seen from Figure 2 EWMA estimate differs greatly with actual value.Also, due to the missing of feedback information,Can be
Remained unchanged in some time, and after hundreds of milliseconds, server waiting list length may occur to change on a large scale.
To solve the timeliness sex chromosome mosaicism, we have carried out simulation reappearance to C3, find when ageing poor, waiting list
Certain law be present in the variation tendency of length.First, although the waiting list length change of server is frequent, its Long-term change trend
It is inviolent, it might even be possible to maintain thousands of milliseconds not change.Secondly, although feedback queue length can not represent current server
State, but it can still reflect the variation tendency of waiting list length.As shown in figure 3,Represent the clothes selected at random
The feedback waiting list length of business device.Found based on more than, the present invention proposes PRS algorithms, in τwDuring more than 100 milliseconds, profit
Server waiting list length is predicted with feedback information, more accurately estimates server queues' situation to obtain.
Prediction of the described PRS algorithms to server end waiting list length by judgement waiting list length change trend and
Predict that two parts of waiting list length form.
Judge that waiting list length change trend is made up of following three steps：
Step 1：By the waiting list length of serverThe service speed μ of corresponding key assignmentssAnd waiting list length
Variation delta qs(being calculated once per 20ms) feeds back to client with return value.
Step 2：According to the waiting list length and Δ q of feedbacksJudge the current variation tendency of waiting list length.
Fig. 4 is waiting list length trend state and its switch condition.The trend of waiting list length is divided into three kinds by PRS
State, steadily, raising and lowering.Judge that the operating process which kind of state is current waiting list length be in is as follows：
Original state is arranged to plateau.PRS enters or 1. the unique conditional of held stationary state is,
Feed back waiting list length and be less than or equal to 30.As long as meeting the condition, PRS immediately enters plateau.
2. 3. the condition of propradation is entered for PRS.If PRS is to be converted to propradation from plateau, say
Bright waiting list length rapid development, therefore switch condition is Δ qs＞ 0 with (In expression
One Δ q fed backs).If it from State Transferring is declined be propradation that PRS, which is, the growth of waiting list length or wait
Treating that the reduction of speed of queue length slows down can show that the variation tendency of waiting list length will turn to be reduced to liter, therefore switch condition
For Δ qs＞ 0 or
4. 5. the condition of decline state is entered for PRS.The design of both conditions is 2. and 3. similar with above-mentioned condition.
In fact PRS can not possibly transit directly to decline state from plateau, but propradation might have feedback information missing,
Therefore, 5. condition is also necessary.
Step 3：Feedback information is selected, only retains the feedback information in same trend and is used to predict waiting list
Length, the feedback information in a upper different variation tendencies preserved before are then deleted.Wherein, in transform boundary
Feedback information then retains if available information, otherwise, it is deleted, it is necessary to judge it.Specifically determination methods are
Subtracted using the waiting list length in a upper feedback information in new feedback information and wait queue length, if result is
Just, and current waiting list length change trend is declines, and the feedback information needs to retain；If result is negative, and current etc.
Queue length variation tendency is treated to rise, the feedback information is also required to retain, otherwise do not retained.
The Boundary Feedback information of three kinds of situations shown in Fig. 5 is available, the information that need to retain.Only have the feedback can be with
During starting point as next trend, judge that the feedback information is useful.
Prediction waiting list length comprises the steps of：
Step 1：Judge the ageing of feedback information, i.e. τwWhether more than 100ms.
Step 2：If τw≤ 100ms, that is, it is strong ageing to think that feedback information has, estimates to wait team using EWMA methods
Row length.The waiting list length in this strong ageing feedback information is taken, it is medium with the feedback informations received all before
Treat that queue length carries out the smooth historical record result of exponential weighting and carries out that exponential weighting is smooth again, using as server end etc.
Treat that queue length is estimated, and result is stored as the smooth historical record result of new exponential weighting, then perform copy sequence
Operation.The historical record result that EWMA methods are taken is performed herein, is referred to since being run PRS, when performing the step, to anti-
After waiting queue length progress exponential weighting smooth in feedforward information, a smooth result of exponential weighting can be obtained, this is tied
Fruit preserves, and the one of input smooth as exponential weighting when running the step next time, and in new feedback information
Waiting list length perform again exponential weighting it is smooth after, new result is covered to original historical record result, so circulation
Operation.
Step 3：If τw＞ 100ms, that is, it is weak ageing to think that feedback information has, using judge waiting list length become
The information that the operation of change trend is collected, carries out waiting list length prediction.Concrete operations are as follows：
If the variation tendency of current waiting list length is steady, PRS, which is used, judges waiting list length change trend
The EWMA values for the feedback information collected are operated as prediction waiting list length value.Calculation formula is as follows：
WhereinBy step 3 in the same waiting list length change trend stored at present, put down using exponential weighting
Sliding algorithm is predicted the historical record of waiting list length.Historical record adopted here, become in waiting list length
In the step of gesture judges three, queue length is waited in the feedback information in the same variation tendency of storage, before use
Waiting list length in other variation tendencies.
If the variation tendency of current waiting list length is rises or declined, PRS, which is used, judges waiting list length
The information that variation tendency operation is collected carries out least square fitting, and substitution current time calculates prediction and waited in fitting function
Queue length value.
In addition, in propradation, if same fitting function has been reused 4 times, it was demonstrated that feedback information is very
Do not update long, in this case, PRS adds an added machinery.(if client has been sent the OSK of this 4 times predictions
But the number of request of feedback is not received) it is equal to 0, then when the 5th is given a mark, directly prediction waiting list length value is set to 0, with true
Protect the server and can determine and be selected to.There is no key assignments request to send past server, it is necessary to attempt for a long-time
Request is sent to obtain its feedback information.Declining state, if predicted value is less than 0, directly set to 0.
Performance Evaluation is carried out to server copy according to the estimation waiting list length that step 2 or step 3 calculate, finally selected
Take the replica server that performance evaluation result is best.The waiting list length value for specifically calculating step 2 or 3 substitutes into marking letter
Number, is ranked up according to the order of score from small to large to copy, chooses the minimum replica server of fraction.Scoring functions are as follows
It is shown：
Wherein, RsThe response time collected for client,For the EWMA values of the service time of feedback information, n is client
Number is held, OSK is the client corresponding server s key assignments access request number for having sent but not responded.
In order to verify the performance of the present invention, emulation experiment has been carried out to PRS and C3.
In simulation flowchart, Poisson distribution is obeyed in the arrival of terminal request.The service time clothes of each key assignments request
From the exponential distribution that average is 4.For the performance inconsistency of emulating server, every the service speed of 500 milliseconds of each servers
The probability for having 50% fluctuates between mean μ and μ * 3.The network delay of one way is 250 microseconds.System shares 200 and please sought survival
Grow up to be a useful person, 50 servers, acquiescence there are 150 clients.Experiment every time repeats to average for five times.Key assignments request sum be
600000。
Fig. 6 is PRS prediction waiting list length and the design sketch of waiting list length change Trend judgement.qpTo make
The waiting list length predicted with PRS methods, qeThe EWMA waiting list length estimates used for C3, this figure show only τw＞
100ms situation.Obviously, PRS prediction waiting list length is more accurate, especially becomes in the change of waiting list length
When gesture is unchanged.The correctness of the judgement of the variation tendency of waiting list length directly influences waiting list length prediction
Order of accuarcy.It can also be seen that the prophetic vision that PRS judges for the variation tendency of waiting list length from Fig. 6.Three kinds
Different line styles represents that waiting list length will enter corresponding three kinds of states.Fig. 6 shows that most anticipation is correct
, and least a portion of erroneous judgement, it can feed back in next time and timely be corrected when coming.In order to more convincing, present invention statistics
PRS and EWMA estimate and the difference of true waiting list length value, as a result show, PRS waits than C3 closer to true
The prediction result of queue length accounts for 57%, and both similar results account for 13%, and remaining 30%C3 approaches compared with PRS.Wherein, two kinds
The difference of method is less than or equal to 1 situation, thinks that the two is similar in statistics.Generally speaking, compared to C3, PRS prediction etc.
Treat that queue length is more accurate
Fig. 7 illustrates PRS algorithms and the survey of 99th% of the C3 algorithms under different clients number environment delay performance
Try design sketch.The change of number clients purpose can directly affect the degree of concurrence of key assignments request and the density of feedback information, because
This client is more, and delay is bigger.It can be seen from figure 7 that PRS algorithms are an advantage over C3 algorithms.
Fig. 8 and Fig. 9 illustrates 99th% of the PRS algorithms with C3 algorithms in the case where key assignments asks deflection delay, point
Not Wei 20% client generation 80% key assignments request and 50% client generation 80% key assignments request.Obviously
Ground, PRS delay are much smaller than C3.In the case where key assignments asks deflection, feedback information is more intensive, and PRS performance is more excellent.
Figure 10 illustrates the influence of delay of the change of copy number to 99th%.Number of copies is increased successively from 3 herein
It is added to 5.Due to the increase of copy number, most fast replica server is selected to become increasingly difficult to, tail delay is also with number of copies increase
And increase.It can be seen from fig. 10 that in number of copies increase, PRS is still better than C3 in terms of tail delay is reduced.
Claims (7)
1. a kind of key assignments data storage storehouse copy selection method based on prediction, it is characterised in that including the wait performed successively
Queue length variation tendency judges, predicts waiting list length and copy three steps of sequence；
Wherein waiting list length Trend judgement includes following three steps：
Step 1：By the letter including waiting list length change amount by way of the return value carrying that key assignments accesses operation
Breath feeds back to client, when client receives feedback information, performs step 2；
Step 2：The variation tendency of the waiting list length of current copy server is judged according to feedback information；
Step 3：If waiting list length change trend does not change, current feedback information and its timestamp are stored, if waited
Queue length variation tendency changes, and stores current feedback information and its timestamp, while deletes a waiting list length
Spend the feedback information of all storages in variation tendency；
Predictive server waiting list length is divided into following three steps：
Step 1：The ageing power of the feedback information newly stored in judgment step three, if strong ageing execution step 2,
If weak ageing execution step 3；
Step 2：The waiting list length in this strong ageing feedback information is taken, and in the feedback informations received all before
Waiting list length carries out the smooth historical record result of exponential weighting, and to carry out exponential weighting again smooth, to be used as server end
Waiting list length estimate, and result is stored as the smooth historical record result of new exponential weighting, then perform copy row
Sequence operates；
Step 3：According to the waiting list length change trend and step 3 of step 2 store information, to server end etc.
Treat that queue length is predicted, perform copy sorting operation；
Copy sorting operation：The estimation waiting list length calculated according to step 2 or step 3 carries out performance to server copy and commented
Estimate, finally choose the best replica server of performance evaluation result.
2. according to the method for claim 1, it is characterised in that in step 1, feedback information includes the wait team in per 5ms
The variation delta q of row lengths, waiting list lengthAnd the service time μ that current key assignments accessess。
3. according to the method for claim 1, it is characterised in that in step 2, according to the variation delta q of waiting list lengths
With waiting list lengthThe variation tendency of waiting list length is divided into three kinds of states, steady, raising and lowering；Initial shape
State is arranged to plateau, and the Rule of judgment changed between three kinds of states is as follows：In any case ifThen judge
For plateau；Δ q need to be met by being converted to propradation from plateaus＞ 0 withSimultaneouslyΔ q need to be met by being converted to decline state from plateaus＜ 0 withSimultaneouslyΔ q need to be met by being converted to decline state from propradations＜ 0 orSimultaneouslyIt is that propradation need to meet Δ q from State Transferring is declineds＞ 0 orSimultaneouslyWhereinRepresent the Δ fed back a qs。
4. according to the method for claim 1, it is characterised in that in step 3, if waiting list length change trend is sent out
Changing, then judge whether a upper feedback information retains by following steps：Use the wait in a upper feedback information
Queue length subtracts in new feedback information and waits queue length, if result is just, and current waiting list length change becomes
To decline, the feedback information needs to retain gesture；If result is negative, and current waiting list length change trend is somebody's turn to do to rise
Feedback information is also required to retain, and does not otherwise retain.
5. according to the method for claim 1, it is characterised in that in step 1, it is to use to judge the ageing power of feedback information
The timestamp of feedback information compares with current time, and threshold value is set as 100ms, has if less than equal to 100ms, feedback information
Strong ageing, if greater than 100ms, feedback information has weak ageing.
6. according to the method for claim 1, it is characterised in that in step 3, if waiting list length change trend is rising
Or declining state, method used in prediction is least-square fitting approach, according to currently stored feedback waiting list length
And its timestamp, linear function of the waiting list length to the time is fitted, substitutes into current time, result of calculation is as pre-
The waiting list length of survey；
If waiting list length change trend is plateau, exponential weighting smoothing algorithm prediction waiting list length is usedMeter
It is as follows to calculate formula：
<mrow>
<mover>
<msup>
<mi>q</mi>
<mo>&prime;</mo>
</msup>
<mo>^</mo>
</mover>
<mo>=</mo>
<mn>0.5</mn>
<mo>*</mo>
<mover>
<mi>q</mi>
<mo>^</mo>
</mover>
<mo>+</mo>
<mn>0.5</mn>
<mo>*</mo>
<msubsup>
<mi>Q</mi>
<mi>S</mi>
<mi>f</mi>
</msubsup>
</mrow>
WhereinBy step 3 in the same waiting list length change trend stored at present, smoothly calculated using exponential weighting
Method is predicted the historical record of waiting list length.
7. according to the method for claim 1, it is characterised in that in copy sorting operation, the operation that is ranked up to copy
It is divided into following steps：The waiting list length value that step 2 or 3 are calculated substitutes into scoring functions, according to score from small to large suitable
Ordered pair copy is ranked up, and chooses the minimum replica server of fraction, and scoring functions are as follows：
Wherein, RsThe response time collected for client,For the EWMA values of the service time of feedback information, n is number clients
Mesh, OSK are the client corresponding server s key assignments access request number for having sent but not responded,It is step 3 in current institute
In the same waiting list length change trend of storage, going through for waiting list length is predicted using exponential weighting smoothing algorithm
The Records of the Historian is recorded.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710631076.8A CN107241442B (en) | 2017-07-28 | 2017-07-28 | A kind of key assignments data storage storehouse copy selection method based on prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710631076.8A CN107241442B (en) | 2017-07-28 | 2017-07-28 | A kind of key assignments data storage storehouse copy selection method based on prediction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107241442A CN107241442A (en) | 2017-10-10 |
| CN107241442B true CN107241442B (en) | 2018-02-09 |
Family
ID=59989803
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710631076.8A Active CN107241442B (en) | 2017-07-28 | 2017-07-28 | A kind of key assignments data storage storehouse copy selection method based on prediction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107241442B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108900509B (en) * | 2018-06-29 | 2020-06-02 | 华中科技大学 | Copy selector based on programmable network equipment |
| CN111444183B (en) * | 2020-03-25 | 2022-08-16 | 中南大学 | Distributed self-adaptive user request scheduling method in key value storage system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103051615A (en) * | 2012-12-14 | 2013-04-17 | 陈晶 | Dynamic defense system capable of resisting large flow attack in honey farm system |
| CN104854831A (en) * | 2012-12-07 | 2015-08-19 | 思科技术公司 | Output queue latency behavior for input queue based device |
| CN105933369A (en) * | 2015-12-24 | 2016-09-07 | ***股份有限公司 | Message forwarding method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9419900B2 (en) * | 2013-12-31 | 2016-08-16 | International Business Machines Corporation | Multi-bit indicator set according to feedback based on an equilibrium length of a queue |
-
2017
- 2017-07-28 CN CN201710631076.8A patent/CN107241442B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104854831A (en) * | 2012-12-07 | 2015-08-19 | 思科技术公司 | Output queue latency behavior for input queue based device |
| CN103051615A (en) * | 2012-12-14 | 2013-04-17 | 陈晶 | Dynamic defense system capable of resisting large flow attack in honey farm system |
| CN105933369A (en) * | 2015-12-24 | 2016-09-07 | ***股份有限公司 | Message forwarding method and device |
Non-Patent Citations (1)
| Title |
|---|
| 基于键值存储的事务控制策略;白皓;《计算机与现代化》;20140228(第2期);第59-62页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107241442A (en) | 2017-10-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11809374B1 (en) | Systems and methods for automatically organizing files and folders | |
| CN106454437B (en) | A kind of streaming media service rate prediction method and device | |
| US20220053341A1 (en) | Transfer learning of network traffic prediction model among cellular base stations | |
| CN108363643B (en) | HDFS copy management method based on file access heat | |
| WO2019050952A1 (en) | Systems, methods, and media for distributing database queries across a metered virtual network | |
| CN106648456B (en) | Dynamic copies file access method based on user's amount of access and forecasting mechanism | |
| CN101226542B (en) | Method for caching report | |
| CN107404409A (en) | Towards the container cloud elastic supply number of containers Forecasting Methodology and system of mutation load | |
| CN110582064B (en) | Short message distribution method, device, equipment and medium | |
| CN103595805A (en) | Data placement method based on distributed cluster | |
| CN107241442B (en) | A kind of key assignments data storage storehouse copy selection method based on prediction | |
| CN111026553A (en) | Resource scheduling method for offline mixed part operation and server system | |
| CN108881432A (en) | Cloud computing cluster load dispatching method based on GA algorithm | |
| EP3557418B1 (en) | Resource management of resource-controlled system | |
| CN109144719A (en) | Cooperation discharging method based on markov decision process in mobile cloud computing system | |
| CN102480502B (en) | I/O load equilibrium method and I/O server | |
| CN113760553A (en) | Mixed cluster task scheduling method based on Monte Carlo tree search | |
| CN113342418B (en) | Distributed machine learning task unloading method based on block chain | |
| CN110266611B (en) | Method, device and system for processing buffered data | |
| US10803036B2 (en) | Non-transitory computer-readable storage medium, data distribution method, and data distribution device | |
| CN116910345A (en) | Label recommending method, device, equipment and storage medium | |
| CN109800059A (en) | A kind of cloud computing virtual machine migration method based on load curve similarity | |
| CN108848514B (en) | Data communication optimization method and data communication optimizer | |
| CN106790485B (en) | Online service request scheduling method based on cost consideration in hybrid cloud mode | |
| CN109062694A (en) | A method of application program is moved into cloud platform |
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 |