CN103152389A - Method and system of responding peak access in cloud computer system - Google Patents
Method and system of responding peak access in cloud computer system Download PDFInfo
- Publication number
- CN103152389A CN103152389A CN2013100413067A CN201310041306A CN103152389A CN 103152389 A CN103152389 A CN 103152389A CN 2013100413067 A CN2013100413067 A CN 2013100413067A CN 201310041306 A CN201310041306 A CN 201310041306A CN 103152389 A CN103152389 A CN 103152389A
- Authority
- CN
- China
- Prior art keywords
- access
- cloud computing
- computing machine
- data
- virtual machine
- 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.)
- Pending
Links
Images
Abstract
The invention belongs to the field of cloud computer system application, and particularly discloses a method of responding peak access in a cloud computer system. The method of responding the peak access in the cloud computer system comprises the following steps: intercepting all access request data at the position of an access point of the cloud computer system, and recording operating data of an operative virtual machine example; predicting access amount of an upcoming predicted point of the cloud computer system according to the access request data; sending an order of building an virtual machine when the predicted access amount is larger than a set value, and selecting a virtual machine example with minimum resource occupation to serve as a target virtual machine used for processing requests, and returning the virtual machine example to the access point so as to carry out processing according to the operating data of the operative virtual machine example. By active load prediction of an access peak, the cloud computer system can start enough resources such as virtual machines to carry out load balancing before the access peak arrives, and really achieves intellectualization and automation of resource scheduling.
Description
Technical field
The invention belongs to the application of cloud computing machine system, specifically, relate to a kind of method and device that is applied to reply peak access in cloud computing machine system.
Background technology
Cloud computing just has been subject to the strong interest of Global Academy, industrial quarters, and has surpassed rapidly grid computing since 2007 are suggested.About the unofficial definition of cloud computing hundreds of, at present comparatively popular comprising is following several.
Wikipedia is defined as cloud computing: " cloud computing is a kind of account form of Internet-based, and in this way, the software and hardware resources of sharing and information can offer computer and other equipment as required.A kind of new IT service increase, use and delivery mode of Internet-based described in cloud computing, and being usually directed to provides dynamically easily expansion and be often virtualized resource by the Internet.Typical cloud computing provider often provides general Network to use, and can visit by the software such as browser or other Web services, and software and data all is stored on server.The key element of cloud computing key comprises that also the Extraordinary user experiences ".
The definition that IBM makes cloud computing in the Technical White Paper for ××× of issue in 2007: " cloud computing is a term that is used for describing certain platform or certain application.Cloud computing platform Dynamical Deployment (provision), configuration (configuration) as required, reconfigure (reconfigure) and cancel deployment server.The server of cloud computing platform can be physics or virtual server.Senior cloud computing platform can comprise other computational resource, as storage area network, the network equipment, fire compartment wall and other safety means.Cloud computing has been described some equally can be by the extendible application program that is accessed to.These cloud computing application programs are come support applications and service with large-scale data center and powerful server.Any equipment that has interface and a standard browser is addressable cloud computing application program all.
Comprehensive above definition, can roughly be summarised as: cloud computing is the product that the traditional calculations machine technology such as grid computing, Distributed Calculation, parallel computation, the network storage, virtual and load balancing and network technical development merge, can be with resource consolidations such as huge system, servers together, so that all kinds of services of high-performance, strong computing capability to be provided; A kind of brand-new computation schema and computing architecture.
Cloud computing is the important component part of generation information technical industry, is the tide of information technology for the third time after personal computer, the Internet, with the essence change of solicited message industry business model.Along with the development of cloud computing and ripe gradually, various novel application begin to introduce cloud computing service, and the construction method of website will turn to from single exploitation, deployment, server or hosting and take the website as core application, construction that have the cloud computing system of distributed environment and outstanding dispatching and load balancing.
As the cloud computing framework take the website cloud as application core, must have distributed scheduling scheme, resource dynamic distribution capability and stable scheduling mechanism.As shown in Figure 1, the core of website cloud scheduling is the scheduling virtual machine device, downward operation to various resources all realizes by virtual machine manager, use unified management interface and data interchange format, the main state collection, data statistics, monitoring etc. that realize resource, all available data upwards are provided, carry out relevant treatment by Service Processing Unit, and carry out alternately with web console.Generally speaking, in order to improve the utilance of virtual machine, the state of all virtual machines that moving of virtual machine manager Real Time Monitoring, and based on the condition of request to the resource consumption maximum of all movable virtual machines, select up till now the virtual machine of most suitable processing process the access request of porch.Yet, although use the scheduling strategy of the utilance that improves virtual machine can effectively improve the utilance of virtual machine, but the peak that enters when cloud computing machine system during the visit, and this strategy just is difficult in time reply the access request of terminal 10, the state of " paralysis " occurs.System is difficult to successfully manage the situation that the peak is accessed for the cloud computing machine, and the researcher has also made many fruitful effort, and is documented among many disclosed patent documentations or Patent Application Publication.for example, Chinese patent 200810114731 discloses a kind of " based on the machine group loading forecast control method of flow model ", this is a kind of forecast Control Algorithm that relates in the complicated production manufacture process that there are many group machine groups in before and after have twice bottleneck operation and per pass bottleneck operation each machine group load of bottleneck operation of rear road, it comprises the timing sampling of machine group load in rear road bottleneck operation, rear road bottleneck operation machine group load desired value is determined, bottleneck operation machine group load d rank, rear road predictive control model is set up and front road bottleneck operation machine group control parameter is asked for.This invention is set up each machine group load estimation control model of rear road bottleneck operation based on flow model and Adaptive Neuro-fuzzy Inference, and the quadratic sum minimum of the difference of road each machine group actual loading of bottleneck operation later on and expectation load is the optimal control target, adopt the Lagrange relaxation method, provide the task output rating of front road each machine group of bottleneck operation, to improve production performance.
In addition, Chinese patent 200810038367 discloses " a kind of grid resource management system and management method ", and this invention has disclosed a kind of grid resource management system, and this system comprises the application layer of grid top layer, the monitoring and control management software layer of grid bottom; Described system also comprises Web service layer, access middleware layer; Application layer, Web service layer, access middleware layer, monitoring and control management software layer connect successively; The Web service layer is used for providing grid resource discovery service, the service of gridding resource historic state, reaches the grid resource scheduling service; Corresponding grid resource discovery module, gridding resource historic state module, and the grid resource scheduling module of comprising; The access middleware layer provides the resource management interface with platform independence, is used for realizing the data access of grid bottom monitoring and control management software layer and Web service layer.This invention grid resource management system can adapt to discovery and the management work of a large amount of dynamic resources in traffic grid by unified management interface is set.
In addition, a kind of client active balancing method of loads in the face of the network multimedia transmission service is also disclosed in Chinese patent CN101184031A, the method is according to the proxy service node number of client connection and the transmission rate of each proxy service node, the number of active accommodation Connection Proxy service node, when the transmission rate of client causes the bandwidth consumption waste over certain threshold value, initiatively reduce transmission node.This active balancing method of loads can be used as replenishing of server end retarded type or forecasting type balancing method of loads, and the response to the Internet Transmission state of raising system improves the multi-media network transmission service quality.
In addition, cloud CDN(Content Delivery Network based on statistical forecast is also disclosed in Chinese patent CN102801792A) the resource automatic deployment method, it is according to 24 hours load estimation values of historical data predict future of each fringe node load of cloud CDN, formulate the virtual server resource deployment plan of respective edges node, each fringe node of cloud CDN is carried out the virtual server resource deployment; Each fringe node loading condition of Real Time Monitoring cloud CDN, when cloud CDN was in the user and accesses the peak period, the mirror image that the cloud platform will configure related service was mounted in virtual server and starts and joins in the middle of each fringe node of cloud CDN, shares the access pressure of burst; When cloud CDN is in non-access peak period, only keep the virtual server of keeping the operation of business normal level, idle virtual server resource is recovered in resource pool.The method can not only successfully manage the peak traffic of burst, improves the resource utilization of CDN, and has reduced energy consumption and the O﹠M cost of CDN.
Yet those skilled in the art still expect can be applied to more method and the device of reply peak access in cloud computing machine system, to solve one or more above-mentioned technical problems.
Summary of the invention
For the deficiencies in the prior art, the objective of the invention is to provide a kind of method and apparatus that is applied to reply peak access in cloud computing machine system, expect that the method and device can improve cloud computing machine system in the response speed of peak Access status.
For this reason, first aspect present invention provides a kind of method that is applied to reply peak access in cloud computing machine system, comprises the steps:
All access request data are tackled in porch in described cloud computing machine system, and the service data of the virtual machine instance of record activity;
Visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
If the visit capacity of future position is greater than or equal to a set point, sends the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.
Correspondingly, second aspect present invention provides a kind of device that is applied to reply peak access in cloud computing machine system, comprising:
Service register is used for tackling all access request data in the porch of described cloud computing machine system, and the service data of the virtual machine instance of record activity;
The machine learning fallout predictor is used for the visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
Load equalizer, be used for after the visit capacity of the future position on the horizon that receives described machine learning fallout predictor surpasses a set point, send the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.
Arbitrary technical scheme of either side of the present invention can make up with the other technologies scheme, as long as they contradiction can not occur.In addition, in arbitrary technical scheme of either side of the present invention, arbitrary technical characterictic goes for this technical characterictic in other embodiments, as long as they contradiction can not occur.Below the invention will be further described.
All documents that the present invention quotes from, their full content is incorporated this paper by reference into, and, if when the implication that these documents are explained and the present invention are inconsistent, be as the criterion with statement of the present invention.In addition, various terms and phrase that the present invention uses have the general sense of well known to a person skilled in the art, nonetheless, the present invention still wishes at this, these terms and phrase to be described in more detail and to explain, the term of mentioning and phrase are as the criterion with the implication that the present invention was explained if any inconsistent with known implication.
Concrete, according to a kind of method that is applied to reply peak access in cloud computing machine system that first aspect present invention provides, it comprises the steps:
All access request data are tackled in porch in described cloud computing machine system, and the service data of the virtual machine instance of record activity;
Visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
If the visit capacity of described future position is greater than or equal to a set point, sends the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.Wherein, described " set point " can be set according to the visit capacity of the daily cardinal principle of described cloud computing machine system, and/or sets according to cloud computing machine system of systems capacity.For example, if the visit capacity of the daily cardinal principle of cloud computing machine system is 100, can set above-mentioned " set point " is 120, thinks that namely predicting visit capacity is to think that the access peak at hand at 120 o'clock.In addition, described " being about to " refers to the moment point that not yet arrives, the moment point that those skilled in the art can Set arbitrarily need to predict, and to improve the ability that cloud computing machine system should the peak access, this paper is not restricted this.
The method of reply peak access in the cloud computing machine that the is applied to system that provides according to first aspect present invention, predict that the visit capacity of the future position on the horizon of described cloud computing machine system comprises the following steps:
Choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n;
Predict the visit capacity v of the future position on the horizon of described cloud computing machine system
knn, wherein:
Wherein, d
iBe predicted value v
n+1To v
iDistance: d
i=|| n – i+1||.
The method of reply peak access in the cloud computing machine that the is applied to system that provides according to first aspect present invention, use the adjacent node algorithm of cum rights K-to comprise the following steps according to the visit capacity of the future position on the horizon of the described cloud computing machine of described forecast sample prediction system:
Choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n, wherein, i is less than or equal to n more than or equal to 1, and, i, n is integer;
Choose the concentrated data of described sample training as the class center C, according to the concentrated data v of described sample training
iTo described class center C apart from d
iWeight w corresponding to these data is set
i, wherein
w
i=f(v
i)=d
i||v
i–v
1||=(i-1)
2||v
i–C||；
By described sample training collection { v
1, v
2, v
3..., v
i..., v
nAnd weight w
iPredict the access number v of next time point
knnThereby, the access peak of completing the described cloud computing machine of prediction system, wherein:
Use the adjacent node algorithm of K-to predict the access peak of described cloud computing machine system by described sample training collection and weights.The adjacent node algorithm of described K-for fixing sample number, meet the weights proportion of target function, the time of its consumption and space complexity be O (1) just obviously, and computational process is extremely quick, the drain space resource, do not implement simple possible.
The method of reply peak access in the cloud computing machine that the is applied to system that provides according to first aspect present invention, the described sampling period is 24 hours, described sample training collection is visit data corresponding to same time point in continuous several sampling periods.
The method of reply peak access in the cloud computing machine that the is applied to system that provides according to first aspect present invention, described class center is visit data corresponding to time point in first sampling period, the data that described sample training is concentrated are far away to the distance at described class center, and weights corresponding to these data are less.
The method of reply peak access in the cloud computing machine that the is applied to system that provides according to first aspect present invention, described weights are selected according to minimum distance classification.About the minimum distance classification method, refer to obtain unknown categorization vector to the distance that will identify representation vector central point of all categories, unknown categorization vector is belonged to a kind of image classification method of the minimum class of distance.those skilled in the art can be with reference to " based on the parameter selection method of the minimum distance classification of core " of Qiu Xiao treasure and Zhang Huaxiang, computer engineering, 05 phase in 2008, do as, once " based on the Classification Method of Itself Adjustment of minimum range principle " of constitution Gui, Institutes Of Technology Of Jiangxi's journal, 04 phase in 2007, Zhang Daoqiang, " the Learning the Kernel Parameters in Kernel Minimum Distance Classifier " of Chen Songcan etc., Pattern Recognition, 2006, 39(1): the documents such as 133-135, with the more deep the present invention that understands.
In addition, according to a kind of device that is applied to reply peak access in cloud computing machine system that second aspect present invention provides, it comprises:
Service register is used for tackling all access request data in the porch of described cloud computing machine system, and the service data of the virtual machine instance of record activity;
The machine learning fallout predictor is used for the visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
Load equalizer, be used for after the visit capacity of the future position on the horizon that receives described machine learning fallout predictor surpasses a set point, send the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.
The device of reply peak access in the cloud computing machine that the is applied to system that provides according to second aspect present invention, described machine learning fallout predictor use the adjacent node algorithm of K-according to the visit capacity of the future position on the horizon of the described cloud computing machine of described forecast sample prediction system.
The device of reply peak access in the cloud computing machine that the is applied to system that provides according to second aspect present invention, described machine learning fallout predictor comprises with lower module:
The first module is used for choosing access request data corresponding to the time point in several sampling periods as the sample training collection;
The second module is used for choosing the concentrated data of described sample training as the class center, and the data concentrated according to described sample training to the distance at described class center arranges weights corresponding to these data;
Computing module be used for to use the adjacent node algorithm of K-to predict the access peak of described cloud computing machine system by described sample training collection and weights.
The device of reply peak access in the cloud computing machine that the is applied to system that provides according to second aspect present invention, the described sampling period is 24 hours, described sample training collection is visit data corresponding to same time point in continuous several sampling periods.
The present invention has obtained following beneficial effect:
The present invention is by the active load estimation to the access peak, cloud computing machine system can open enough resources (for example virtual machine) and carry out load balancing before the peak access arrives, transfer aggressive mode to from Passive Mode, break away from the situations such as the response lag that system produces, stability deficiency, really realized intellectuality and the automation of scheduling of resource.
Description of drawings
Accompanying drawing is not to be intended to draw in proportion.In these accompanying drawings, pass through same reference numeral at each the identical or almost identical parts shown in each accompanying drawing.In order to know purpose, in the accompanying drawings each parts is not carried out mark.In the accompanying drawings:
Fig. 1 is the existing scheduling of resource schematic diagram that is applied to cloud computing machine system, can find out in figure, this cloud computing machine system utilizes the running status of virtual machine manager Real Time Monitoring virtual machine, selects at present the virtual machine of the most suitable processing to process the access request of porch, improves the utilance of virtual machine.
Fig. 2 is the flow chart of an embodiment of method of the present invention, can find out in figure, by can in time adjust the resource deployment of cloud computing machine system to the prediction of visit capacity, improves cloud computing machine system in the response speed of access peak period;
Fig. 3 is the flow chart of an embodiment of the visit capacity step of prediction cloud computing machine system in Fig. 2, utilizes the adjacent node algorithm of K-to realize the prediction on access peak in this embodiment;
Fig. 4 is the flow chart of an embodiment of the visit capacity step of prediction cloud computing machine system in Fig. 2, utilizes the adjacent node algorithm of cum rights K-to realize the prediction on access peak in this embodiment;
Fig. 5 is of the present invention for the flow chart of cum rights KNN method;
Fig. 6 is the structural representation of an embodiment of device of the present invention, can find out in figure, by can in time adjust the resource deployment of cloud computing machine system to the prediction of visit capacity, improves cloud computing machine system in the response speed of access peak period;
Fig. 7 is the structural representation of machine learning fallout predictor in Fig. 6, by to the choosing of the weights of sample training collection and each data, can use the adjacent node algorithm of cum rights K-to realize rapidly the prediction of visit capacity;
Fig. 8 is based on the illustraton of model of the cloud computing machine system of Eucalyptus;
Fig. 9 is the block diagram of experimental result, and as can be seen from the figure, machine learning algorithm can initiatively be opened virtual machine and process before the access peak arrives, successfully realized prediction and processing to the peak visit capacity, reached re-set target;
In figure:
10-terminal; 21-entrance; 22-virtual machine manager; 23-virtual machine; 31-service register; 32-load equalizer; 33-machine learning fallout predictor; The 331-the first module; The 332-the second module; 333-computing module.
Embodiment
Further illustrate the present invention below by object lesson, still, should be understood to, these examples are only used for the use that specifically describes more in detail, and should not be construed as for limiting in any form the present invention.
Embodiments of the invention can carry out in any cloud computing machine system with scheduling of resource, load monitoring ability.By to the prediction of the visit capacity of cloud computing machine system and timely adjustment and deployment (can be for example to increase virtual machine), thereby during arriving on the access peak, the raising system improves the responsibility of system.
See also Fig. 2, in one embodiment of the invention, the described method that is applied to reply peak access in cloud computing machine system comprises the steps:
Step S201: all access request data are tackled at entrance 21 places in described cloud computing machine system, and the service data of virtual machine 23 examples of record activity.Wherein, the access request data of interception can be for example the access request of HTTP.The visit data that uses phrase " interception " to refer to the to receive entrance 21 places herein line item of going forward side by side does not affect the response of its access request.
Step S202: according to the visit capacity of the future position on the horizon of the described cloud computing machine of described access request data prediction system.About the prediction to visit capacity, generally by sample set off in search rule, and utilize rule that unknown data is predicted, can be for example SVM(Support Vector Machine, SVMs) algorithm has effectively overcome the shortcomings such as sample distribution, redundancy feature, overfitting, have good generalization ability, take advantage on stability.Yet disappointed is that the convergence speed of SVM on large data sets is slower, needs a large amount of storage resources and powerful calculating ability to become its deadly defect.And the method abilities such as Bayes, linear classification, decision tree and KNN relatively a little less than, but their model is simple, efficient is high.In addition, also can use KNN(K arest neighbors sorting algorithm) method predicts, preferably uses cum rights KNN method to predict.KNN method about KNN method and cum rights will be set forth follow-up, wouldn't be discussed herein.
Step S203: the test access peak whether at hand, if the visit capacity of described future position is greater than or equal to a set point, namely predicting the access peak will arrive at the time point of next one prediction, forward step S204 to, otherwise, return to step S201.At this moment, virtual machine 23 managers 22 still operate in normal mode, can be for example to dispatch as target take the peak use rate of virtual machine 23, therefore can not make intervention to system.In addition, apparent, step S201 is the action of a continuation, no matter subsequent scenario how, all access request data all can be tackled by cloud computing machine system, for next step prediction provides Data support.；
Step S204: if predict the access peak at hand, the running status in the example pond of current virtual machine 23 and select minimum virtual machine 23 examples of resource occupation to return to entrance 21 places as the target virtual machine 23 of processing request according to the service data of virtual machine 23 examples of described activity and process no matter send the instruction that creates virtual machine 23.At this moment, original operation strategy of virtual machine 23 managers 22 will suspend execution, the scheduling meeting that for example realizes high usage is suspended, and selects minimum virtual machine 23 examples of resource occupation to return to entrance 21 with the response access request as the target virtual machine 23 of processing request.
See also Fig. 3, Figure 3 shows that KNN(K arest neighbors sorting algorithm) block diagram.The KNN method is a kind of machine learning algorithm of theoretical comparative maturity, its core concept is: if the great majority in the sample of the K of sample in feature space (being the most contiguous in feature space) the most similar belong to some classifications, this sample also belongs to this classification.For website cloud load estimation strategy, the KNN algorithm all has higher model dependency from the foundation of choosing feature space etc. of sampling, sample training collection.
As shown in Figure 3, use the method for KNN method reply peak access, predict that the visit capacity of the future position on the horizon of described cloud computing machine system comprises the following steps:
Step S201: choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n.The below illustrates how to build access request data as the sample training collection for example.Can be for example as shown in Figure 4, access request to entrance 21 places was tackled and record according to the employing cycle, the continuous sampling period of preferred use is carried out record, for example take one day as a sampling period, carry out sampled data take Monday as first sampling period, carry out sampled data take Tuesday as second sampling period ..., by that analogy.All sample in identical moment point in each sampling period j, obtain as cycle index data Aj1, Aj2 ..., Aji ..., Ajn}, the i.e. record of horizontal data.In these sampling periods, the data that record in the same time mutually, for example A1i, A2i ..., Ami} is as a sample training collection Ri, and namely sample training collection Ri belongs to a longitudinal data, thereby if can obtain each and every one sample training collection Ri.
Step S302: to each described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn choose data as the class center C of this sample training collection, wherein, i is less than or equal to n more than or equal to 1, and, i, n is integer.
Step S303: the visit capacity v that predicts the future position on the horizon of described cloud computing machine system
knn, wherein:
Wherein, d
iBe predicted value v
n+1To v
iDistance: d
i=|| n – i+1||.
According to the difference of the sample training collection of choosing, can predict next visit capacity constantly corresponding to each moment point, thus the prediction of realization access peak value.
In addition, the visit capacity of the described cloud computing machine of prediction system is preferably used another follow-on KNN method.See also Fig. 5, Figure 5 shows that cum rights KNN method, it comprises the following steps:
Step S501: choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n, wherein, i is less than or equal to n more than or equal to 1, and, i, n is integer.The sample training collection that uses in sample training collection used herein and Fig. 3 is consistent, no longer gives unnecessary details herein.
Step S502: choosing data that described sample training concentrates as the class center C, can be for example that described class center is visit data v corresponding to time point in first sampling period
1According to the concentrated data v of described sample training
iTo described class center C apart from d
iWeight w corresponding to these data is set
i, wherein
w
i=f(v
i)=d
i||v
i–v
1||=(i-1)
2||v
i–C||；
Preferably, the data that described sample training is concentrated are far away to the distance of described class center C, and weights corresponding to these data are less, and select according to minimum distance classification.
Step S503: by described sample training collection { v
1, v
2, v
3..., v
i..., v
nAnd weight w
iPredict the access number v of next time point
knnThereby, the access peak of completing the described cloud computing machine of prediction system, wherein:
Corresponding, the present invention also provides the device that is applied to reply peak access in cloud computing machine system, and as shown in Figure 6, it also comprises:
Service register 31 is used for tackling all access request data at entrance 21 places of described cloud computing machine system, and the service data of virtual machine 23 examples of record activity;
Machine learning fallout predictor 33 is used for the visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
Load equalizer 32, be used for after the visit capacity of the future position on the horizon that receives described machine learning fallout predictor 33 surpasses a set point, namely access the peak at hand, send the instruction that creates virtual machine 23 and select minimum virtual machine 23 examples of resource occupation to return to entrance 21 places as the target virtual machine 23 of processing request according to the service data of virtual machine 23 examples of described activity and process.(namely access under the situation that does not arrive on the peak) under normal circumstances, manage, dispatch to respond the access request at entrance 21 places by 22 pairs of virtual machines 23 of virtual machine 23 manager.In this process, service register 31 also can carry out record to the access request at entrance 21 places, and is sent to machine learning fallout predictor 33 and calculates, and predicts the peak of its access request.When machine learning fallout predictor 33 predicts the access peak at hand the time, original operation strategy of virtual machine 23 managers 22 will suspend execution, the scheduling meeting that for example realizes high usage is suspended, and selects minimum virtual machine 23 examples of resource occupation to return to entrance 21 with the response access request as the target virtual machine 23 of processing request.
Accordingly, described machine learning fallout predictor 33 can use the adjacent node algorithm of K-according to the access peak of the described cloud computing machine of described forecast sample prediction system.Use the machine learning fallout predictor 33 of the adjacent node algorithm of K-to comprise following several module:
Module one: be used for choosing access request data corresponding to the time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n.
The below illustrates how to build access request data as the sample training collection for example.Can be for example as shown in Figure 4, access request to entrance 21 places was tackled and record according to the employing cycle, the continuous sampling period of preferred use is carried out record, for example take one day as a sampling period, carry out sampled data take Monday as first sampling period, carry out sampled data take Tuesday as second sampling period ..., by that analogy.All sample in identical moment point in each sampling period j, obtain as cycle index data Aj1, Aj2 ..., Aji ..., Ajn}, the i.e. record of horizontal data.In these sampling periods, the data that record in the same time mutually, for example A1i, A2i ..., Ami} is as a sample training collection Ri, and namely sample training collection Ri belongs to a longitudinal data, thereby if can obtain each and every one sample training collection Ri.
Module two: be used for each described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn choose data as the class center C of this sample training collection, wherein, i is less than or equal to n more than or equal to 1, and, i, n is integer.
Module three: the visit capacity v that predicts the future position on the horizon of described cloud computing machine system
knn, wherein:
Wherein, d
iBe predicted value v
n+1To v
iDistance: d
i=|| n – i+1||.
In addition, described machine learning fallout predictor 33 also can use another structure to predict, as shown in Figure 7, this structure comprises with lower module:
The first module 331 is used for choosing access request data corresponding to the time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n; Choose discuss in module one in the adjacent node algorithm of the process of sample training collection such as above-mentioned K-consistent, no longer give unnecessary details herein.Generally speaking, preferred, the described sampling period is 24 hours, described sample training collection { v
1, v
2, v
3..., v
i..., v
nBe visit data corresponding to same time point in continuous several sampling periods.
The second module 332 is used for choosing described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn data as the class center C, according to described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn data v
iTo described class center C apart from d
iThese data v is set
iCorresponding weight w
i
Illustrate implementation procedure of the present invention and technique effect below by an experiment.
The below describes as embodiment as website cloud basic platform framework with the Eucalyptus system.
See also Fig. 8, Fig. 8 is the website cloud architecture based on Eucalyptus, and on it, virtual machine 23 of operation comprises Web container and Web engineering, is a complete Web server.The Eucalyptus front end node is responsible for scheduling of resource and the control of system, and backend nodes provides the Web server ability based on the website.
System is carried out a continuous test of week, carry out in advance data record in test the first seven day, the data that are saved will be as the sample training collection of machine learning fallout predictor 33.Due to the predictability and the load balancing that are test reply peak access, therefore, by the computing that began to carry out machine learning on the 8th day.Utilize LoadRunner as the pressure test instrument, the concurrent users of simulation varying number carry out peak access experiment to the some websites example in the cloud platform of website, according to quantity and the average utilization of virtual machine 23 under the change records of access number.
As shown in Figure 9, by experimental result as can be known, when visit capacity rises to 1500, the machine learning predicting machine has predicted possible peak access and has newly started a new virtual machine 23 according to the sample training collection, along with the continuous increase of visit capacity and reach at 2500 o'clock, system starts to 4 virtual machines 23 and processes, thereby can in time reply the access of system.Experimental result shows, machine learning algorithm can initiatively be opened virtual machine 23 and process before the official visit peak arrives, and successfully realized prediction and processing that the peak is accessed reaching re-set target.
The above has done detailed description to embodiments of the present invention by reference to the accompanying drawings, but the present invention is not limited to above-mentioned execution mode, in the ken that one skilled in the relevant art possesses, can also make various variations under the prerequisite that does not break away from aim of the present invention.
Claims (10)
1. a method that is applied to reply peak access in cloud computing machine system, is characterized in that, comprises the steps:
All access request data are tackled in porch in described cloud computing machine system, and the service data of the virtual machine instance of record activity;
Visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
If the visit capacity of described future position is greater than or equal to a set point, sends the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.
2. the method that is applied to reply peak access in cloud computing machine system as claimed in claim 1, is characterized in that, predicts that the visit capacity of the future position on the horizon of described cloud computing machine system comprises the following steps:
Choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n;
Predict the visit capacity v of the future position on the horizon of described cloud computing machine system
knn, wherein:
Wherein, d
iBe predicted value v
n+1To v
iDistance: d
i=|| n – i+1||.
3. the method that is applied to reply peak access in cloud computing machine system as claimed in claim 1, is characterized in that, predicts that the visit capacity of the future position on the horizon of described cloud computing machine system comprises the following steps:
Choose access request data corresponding to a time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n, wherein, i is less than or equal to n more than or equal to 1, and, i, n is integer;
Choose the concentrated data of described sample training as the class center C, according to the concentrated data v of described sample training
iTo described class center C apart from d
iWeight w corresponding to these data is set
i, wherein
w
i=f(v
i)=d
i||v
i–v
1||=(i-1)
2||v
i–C||；
By described sample training collection { v
1, v
2, v
3..., v
i..., v
nAnd weight w
iPredict the access number v of next time point
knnThereby, the access peak of completing the described cloud computing machine of prediction system, wherein:
4. be applied to as claimed in claim 2 or claim 3 the method for reply peak access in cloud computing machine system, it is characterized in that, the described sampling period is 24 hours, and described sample training collection is visit data corresponding to same time point in continuous several sampling periods.
5. the method that is applied to reply peak access in cloud computing machine system as claimed in claim 3, it is characterized in that, described class center is visit data corresponding to time point in first sampling period, the data that described sample training is concentrated are far away to the distance at described class center, and weights corresponding to these data are less.
6. the method that is applied to reply peak access in cloud computing machine system as claimed in claim 5, is characterized in that, described weights are selected according to minimum distance classification.
7. a device that is applied to reply peak access in cloud computing machine system, is characterized in that, comprising:
Service register is used for tackling all access request data in the porch of described cloud computing machine system, and the service data of the virtual machine instance of record activity;
The machine learning fallout predictor is used for the visit capacity according to the future position on the horizon of the described cloud computing machine of described access request data prediction system;
Load equalizer, be used for after the visit capacity of the future position on the horizon that receives described machine learning fallout predictor surpasses a set point, send the instruction that creates virtual machine and select the minimum virtual machine instance of resource occupation to return to the porch as the target virtual machine of processing request according to the service data of the virtual machine instance of described activity and process.
8. the device that is applied to reply peak access in cloud computing machine system as claimed in claim 7, it is characterized in that, described machine learning fallout predictor uses the adjacent node algorithm of K-according to the visit capacity of the future position on the horizon of the described cloud computing machine of described forecast sample prediction system.
9. the device that is applied to reply peak access in cloud computing machine system as claimed in claim 7, is characterized in that, described machine learning fallout predictor comprises with lower module:
The first module is used for choosing access request data corresponding to the time point in several sampling periods as sample training collection { v
1, v
2, v
3..., v
i..., v
n;
The second module is used for choosing described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn data as the class center C, according to described sample training collection { v
1, v
2, v
3..., v
i..., v
nIn data v
iTo described class center C apart from d
iThese data v is set
iCorresponding weight w
i
Computing module is used for according to described sample training collection { v
1, v
2, v
3..., v
i..., v
nAnd weight w
iPredict the access number v of next time point
knnThereby, the access peak of completing the described cloud computing machine of prediction system, wherein:
10. be applied to as claimed in claim 8 or 9 the device of reply peak access in cloud computing machine system, it is characterized in that, the described sampling period is 24 hours, described sample training collection { v
1, v
2, v
3..., v
i..., v
nBe visit data corresponding to same time point in continuous several sampling periods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013100413067A CN103152389A (en) | 2013-02-01 | 2013-02-01 | Method and system of responding peak access in cloud computer system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013100413067A CN103152389A (en) | 2013-02-01 | 2013-02-01 | Method and system of responding peak access in cloud computer system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103152389A true CN103152389A (en) | 2013-06-12 |
Family
ID=48550248
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013100413067A Pending CN103152389A (en) | 2013-02-01 | 2013-02-01 | Method and system of responding peak access in cloud computer system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103152389A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103327118A (en) * | 2013-07-09 | 2013-09-25 | 南京大学 | Intelligent virtual machine cluster scaling method and system for web application in cloud computing |
| CN103442076A (en) * | 2013-09-04 | 2013-12-11 | 上海海事大学 | Usability guarantee method for cloud storage system |
| CN104580524A (en) * | 2015-01-30 | 2015-04-29 | 华为技术有限公司 | Resource scaling method and cloud platform with same |
| WO2015077958A1 (en) * | 2013-11-28 | 2015-06-04 | 华为技术有限公司 | Method, apparatus and system for controlling service traffic |
| WO2016065858A1 (en) * | 2014-10-30 | 2016-05-06 | 中兴通讯股份有限公司 | Method and device for deploying virtual core network |
| CN106027685A (en) * | 2016-08-02 | 2016-10-12 | 合肥奇也信息科技有限公司 | Peak access method based on cloud computation system |
| CN106230944A (en) * | 2016-08-02 | 2016-12-14 | 合肥奇也信息科技有限公司 | The running gear that a kind of peak based on cloud computer system accesses |
| CN106330754A (en) * | 2016-08-31 | 2017-01-11 | 东软集团股份有限公司 | Access request control method and device |
| CN108023742A (en) * | 2016-10-31 | 2018-05-11 | 阿里巴巴集团控股有限公司 | A kind of expansion method of application, device and system |
| CN108880945A (en) * | 2018-08-02 | 2018-11-23 | 浙江口碑网络技术有限公司 | A kind of cloud monitoring system and method |
| CN108984344A (en) * | 2018-07-12 | 2018-12-11 | 郑州云海信息技术有限公司 | A kind of dynamic generates method, apparatus, equipment and the storage medium of snapshot |
| CN109104497A (en) * | 2018-09-26 | 2018-12-28 | 郑州云海信息技术有限公司 | A kind of method for processing business and device based on cloud platform |
| US10217053B2 (en) | 2015-06-23 | 2019-02-26 | International Business Machines Corporation | Provisioning service requests in a computer system |
| CN109492774A (en) * | 2018-11-06 | 2019-03-19 | 北京工业大学 | A kind of cloud resource dispatching method based on deep learning |
| CN110493643A (en) * | 2019-08-06 | 2019-11-22 | 北京邮电大学 | Video storage method and device |
| WO2019227707A1 (en) * | 2018-06-01 | 2019-12-05 | 平安科技（深圳）有限公司 | Server, method for predicting system invocation volume, and storage medium |
| CN111738781A (en) * | 2020-07-31 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Deep learning model and method for predicting flow data thereof |
| CN111858015A (en) * | 2019-04-25 | 2020-10-30 | 中国移动通信集团河北有限公司 | Method, device and gateway for configuring running resources of application program |
| CN113014671A (en) * | 2021-04-01 | 2021-06-22 | 毕延杰 | Cloud service resource sharing method and resource sharing platform applied to big data analysis |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6714968B1 (en) * | 2000-02-09 | 2004-03-30 | Mitch Prust | Method and system for seamless access to a remote storage server utilizing multiple access interfaces executing on the remote server |
| CN102427475A (en) * | 2011-12-08 | 2012-04-25 | 曙光信息产业(北京)有限公司 | Load balance scheduling system in cloud computing environment |
-
2013
- 2013-02-01 CN CN2013100413067A patent/CN103152389A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6714968B1 (en) * | 2000-02-09 | 2004-03-30 | Mitch Prust | Method and system for seamless access to a remote storage server utilizing multiple access interfaces executing on the remote server |
| CN102427475A (en) * | 2011-12-08 | 2012-04-25 | 曙光信息产业(北京)有限公司 | Load balance scheduling system in cloud computing environment |
Non-Patent Citations (1)
| Title |
|---|
| 刘子锋: "基于Eucalyptus网站云的资源调度", 《计算机系统应用》, 15 October 2012 (2012-10-15) * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103327118A (en) * | 2013-07-09 | 2013-09-25 | 南京大学 | Intelligent virtual machine cluster scaling method and system for web application in cloud computing |
| CN103327118B (en) * | 2013-07-09 | 2016-10-05 | 南京大学 | A kind of in cloud computing web application intelligent virtual computer cluster Zoom method and system |
| CN103442076B (en) * | 2013-09-04 | 2016-09-07 | 上海海事大学 | A kind of usability guarantee method of cloud storage system |
| CN103442076A (en) * | 2013-09-04 | 2013-12-11 | 上海海事大学 | Usability guarantee method for cloud storage system |
| WO2015077958A1 (en) * | 2013-11-28 | 2015-06-04 | 华为技术有限公司 | Method, apparatus and system for controlling service traffic |
| WO2016065858A1 (en) * | 2014-10-30 | 2016-05-06 | 中兴通讯股份有限公司 | Method and device for deploying virtual core network |
| CN105610592A (en) * | 2014-10-30 | 2016-05-25 | 中兴通讯股份有限公司 | Method and device for deploying virtual core network |
| WO2016119412A1 (en) * | 2015-01-30 | 2016-08-04 | 华为技术有限公司 | Resource scaling method on cloud platform and cloud platform |
| CN104580524A (en) * | 2015-01-30 | 2015-04-29 | 华为技术有限公司 | Resource scaling method and cloud platform with same |
| US10217053B2 (en) | 2015-06-23 | 2019-02-26 | International Business Machines Corporation | Provisioning service requests in a computer system |
| CN106027685A (en) * | 2016-08-02 | 2016-10-12 | 合肥奇也信息科技有限公司 | Peak access method based on cloud computation system |
| CN106230944A (en) * | 2016-08-02 | 2016-12-14 | 合肥奇也信息科技有限公司 | The running gear that a kind of peak based on cloud computer system accesses |
| CN106330754A (en) * | 2016-08-31 | 2017-01-11 | 东软集团股份有限公司 | Access request control method and device |
| CN106330754B (en) * | 2016-08-31 | 2020-02-21 | 东软集团股份有限公司 | Access request control method and device |
| CN108023742A (en) * | 2016-10-31 | 2018-05-11 | 阿里巴巴集团控股有限公司 | A kind of expansion method of application, device and system |
| CN108023742B (en) * | 2016-10-31 | 2021-08-27 | 阿里巴巴集团控股有限公司 | Application capacity expansion method, device and system |
| WO2019227707A1 (en) * | 2018-06-01 | 2019-12-05 | 平安科技（深圳）有限公司 | Server, method for predicting system invocation volume, and storage medium |
| CN108984344A (en) * | 2018-07-12 | 2018-12-11 | 郑州云海信息技术有限公司 | A kind of dynamic generates method, apparatus, equipment and the storage medium of snapshot |
| CN108984344B (en) * | 2018-07-12 | 2021-09-17 | 郑州云海信息技术有限公司 | Method, device and equipment for dynamically generating snapshot and storage medium |
| CN108880945B (en) * | 2018-08-02 | 2020-07-28 | 浙江口碑网络技术有限公司 | Cloud monitoring system and method |
| CN108880945A (en) * | 2018-08-02 | 2018-11-23 | 浙江口碑网络技术有限公司 | A kind of cloud monitoring system and method |
| CN109104497A (en) * | 2018-09-26 | 2018-12-28 | 郑州云海信息技术有限公司 | A kind of method for processing business and device based on cloud platform |
| CN109492774A (en) * | 2018-11-06 | 2019-03-19 | 北京工业大学 | A kind of cloud resource dispatching method based on deep learning |
| CN109492774B (en) * | 2018-11-06 | 2021-10-26 | 北京工业大学 | Deep learning-based cloud resource scheduling method |
| CN111858015A (en) * | 2019-04-25 | 2020-10-30 | 中国移动通信集团河北有限公司 | Method, device and gateway for configuring running resources of application program |
| CN111858015B (en) * | 2019-04-25 | 2024-01-12 | 中国移动通信集团河北有限公司 | Method, device and gateway for configuring running resources of application program |
| CN110493643A (en) * | 2019-08-06 | 2019-11-22 | 北京邮电大学 | Video storage method and device |
| CN111738781A (en) * | 2020-07-31 | 2020-10-02 | 支付宝(杭州)信息技术有限公司 | Deep learning model and method for predicting flow data thereof |
| CN113014671A (en) * | 2021-04-01 | 2021-06-22 | 毕延杰 | Cloud service resource sharing method and resource sharing platform applied to big data analysis |
| CN113014671B (en) * | 2021-04-01 | 2021-11-23 | 湖南机械之家信息科技有限公司 | Cloud service resource sharing method and resource sharing platform applied to big data analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103152389A (en) | Method and system of responding peak access in cloud computer system | |
| Sahoo et al. | ESMLB: Efficient switch migration-based load balancing for multicontroller SDN in IoT | |
| Aujla et al. | MEnSuS: An efficient scheme for energy management with sustainability of cloud data centers in edge–cloud environment | |
| Wang et al. | MAP based modeling method and performance study of a task offloading scheme with time-correlated traffic and VM repair in MEC systems | |
| CN110231976B (en) | Load prediction-based edge computing platform container deployment method and system | |
| CN111930511A (en) | Identifier resolution node load balancing device based on machine learning | |
| Xia | Resource scheduling for piano teaching system of internet of things based on mobile edge computing | |
| Sigwele et al. | Fuzzy logic-based call admission control in 5G cloud radio access networks with preemption | |
| Dimitrios et al. | Simulation and performance evaluation of a fog system | |
| Zare et al. | Towards distributed and autonomous IoT service placement in fog computing using asynchronous advantage actor-critic algorithm | |
| Swain et al. | Efficient resource management in cloud environment | |
| Saxena et al. | Workload forecasting and resource management models based on machine learning for cloud computing environments | |
| Tekiyehband et al. | An efficient dynamic service provisioning mechanism in fog computing environment: A learning automata approach | |
| Zhang | Storage optimization algorithm design of cloud computing edge node based on artificial intelligence technology | |
| Liu et al. | ScaleFlux: Efficient stateful scaling in NFV | |
| Karimiafshar et al. | A request dispatching method for efficient use of renewable energy in fog computing environments | |
| Li et al. | Service cost-based resource optimization and load balancing for edge and cloud environment | |
| Wang et al. | A quick adaptive migration algorithm for virtual network function | |
| Zhao et al. | Application of dynamic management of 5G network slice resource based on reinforcement Learning in Smart Grid | |
| Cui et al. | Resource-Efficient DNN Training and Inference for Heterogeneous Edge Intelligence in 6G | |
| Zhu et al. | Deep reinforcement learning-based edge computing offloading algorithm for software-defined IoT | |
| Surya et al. | Novel Approaches for Resource Management Across Edge Servers | |
| Li et al. | Two-stage selection of distributed data centers based on deep reinforcement learning | |
| Sha et al. | A multi-objective QoS-aware IoT service placement mechanism using Teaching Learning-Based Optimization in the fog computing environment | |
| Tang et al. | A Survey on Scheduling Techniques in Computing and Network Convergence |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C12 | Rejection of a patent application after its publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20130612 |