CN103873569A - Resource optimized deployment method based on IaaS (infrastructure as a service) cloud platform - Google Patents

Abstract

The invention relates to a resource optimized deployment method based on an IaaS (infrastructure as a service) cloud platform. Application loads of different users are classified, so that the IaaS cloud platform can deploy virtual machine (VM) resources and physical resources according to different classifications; therefore the resource utilization rate of the cloud platform is maximized, application features can be extracted and processed, a BP (back propagation) neural network and a classifier model are constructed, and distribution is implemented by a TAEA algorithm. The resource optimized deployment method has the beneficial effects that the classifier model is constructed, and the TAEA algorithm is adopted, so that the deployment priority levels of two factors, namely an application load feature and an SLA (service level agreement), can be judged; therefore which factor deploys the application load of a VM to a divided PM (physical machine) is considered in priority, and the energy consumption of the whole data center is the lowest.

Description

A kind of resource optimization dispositions method based on IaaS cloud platform

Technical field

The invention belongs to cloud platform application technical field, relate in particular to the method for designing that reduces the resource load grader of energy consumption by IaaS cloud computing.

Background technology

In recent years, along with the development of Intel Virtualization Technology, particularly have all many-sided advantages such as Server Consolidation, online migration, isolation, high availability, deployment flexibly, low management cost take Intel Virtualization Technology as basic cloud computing technology, it has become the main direction of Future Data center development.And consider energy-conservation IaaS (Infrastructure as a Service) cloud platform, provide data center, architecture hardware and software resource by the Internet, IT infrastructure ability is offered to user to be used, it has following characteristics: the load of (1) IaaS cloud platform is constantly to change, and IaaS provider is difficult to the Changing Pattern of precognition load; (2) the basic management unit of IaaS cloud platform is virtual machine (Virtual Machine, VM), VM can, the in the situation that of user's unaware, move to virtual machine on another physical host from a physical host (Physical Machine, PM) by online migrating technology.Consider based on above two features and from the energy-conservation visual angle of IaaS cloud platform, can sum up and draw: Server Consolidation technology and virtual machine (vm) migration technology.

No matter be Server Consolidation or virtual machine (vm) migration, it is in fact all that IaaS cloud platform resource is carried out to rational management, make cloud platform can provide the service (SLA(Service Level Agreement) of efficient stable to guarantee), and can reduce energy consumption cost, ensure the maximization of service income.

For Server Consolidation technology, must consider the problem of following two aspects, (1) whether infrastructure supports that software and hardware is virtual simultaneously, and accomplish to optimize resource deployment? (2) whether resource deployment optimization is considered comprehensive, comprise the resources such as CPU, internal memory, disk, network? the elaboration of these two problems, has directly embodied comprehensive that the importance of physical resource utilance and server integration technology consider.

Virtual machine (vm) migration technical elements, migrating technology is commonly referred to as dynamic migration technology, and static migrating technology can not meet current needs.With regard to dynamic migration technology on being applied to IaaS cloud platform, particularly important with which type of migration strategy combination terrain, wherein must consider the problem of following several aspects, can (1) VM migration respond quickly according to the variation of application load? (2) there is the ability of processing multi-user's multiple types SLA application? (3) whether there is certain robustness, can and support variety classes application and mixed load? for the current most of virtual machine (vm) migration of problem (1), technical research institute does not consider, it has embodied the elasticity of IaaS cloud platform energy-saving distribution strategy, this is one of the present patent application problem to be solved just, the virtual machine (vm) migration technology that is IaaS cloud platform for the requirement of problem (2) will be for the adaptive different strategy of different application, and to reach the requirement that adapts to different application type of user, this is also the feature that the present patent application has, there is the ability that can support the large-scale application load of variety classes for problem (3) major requirement VM migrating technology.

Summary of the invention

For the deficiencies in the prior art, the object of the present invention is to provide a kind of resource optimization dispositions method based on IaaS cloud platform, by flexible scheduling strategy and TAEA algorithm, improve the resource deployment ability of IaaS cloud platform.

Technical scheme of the present invention is as follows:

Based on a resource optimization dispositions method for IaaS cloud platform, comprise the following steps:

Step S1: analytical applications load characteristic and SLA level characteristics, and carry out extraction and the pattern recognition of load Sampling characters, its result comprises characteristic and the classification mode that preliminary treatment obtains;

Step S2: extract and create a BP neural net based on load characteristic, its input vector X is described characteristic, and its output vector Y is described classification mode, and the structure of BP neural net is drawn by feature extraction and the pattern matching of system input and output;

Step S3: create sorter model based on described BP neural net, and it is trained and draws each layer of optimum weights and threshold value of sorter model;

Step S4: use TAEA algorithm, the later application load of classifying is dispatched, this application load is deployed on most suitable server, realize the optimum allocation of virtual machine.

Above-mentioned dispositions method, in described step S1, adopts the statistical theories such as IRQ interquartile-range IQR to carry out the extraction of load Sampling characters, to guarantee to embody the feature of original load sampling curve.

Above-mentioned dispositions method, described load Sampling characters at least comprises cpu busy percentage and the memory usage of a period of time with interior application load, and described load Sampling characters is carried out to following operation:

(1) determine the signature identification of frequency, time window and the SLA rank of load sampling;

(2) determine the attribute and the form that extract feature.

Above-mentioned dispositions method, described load Sampling characters is formulated and initialization by the keeper of cloud platform.

Above-mentioned dispositions method, described BP neural net is single hidden layer, it is take resource load number of categories to be sorted as output node number, and the resource load grader modeling based on BP neural net comprises structure, the training of network and three parts of the classification of network of network.

Further, the convergence rate of described sorter model is more than or equal to IaaS cloud platform provides the SLA requirement of service.

Above-mentioned dispositions method, triggers scheduling of resource by the resource allocation algorithm of described sorter model, according to the classification of application load characteristic of correspondence, corresponding application load is deployed on PM effectively.

Above-mentioned dispositions method, in described step S4, described TAEA algorithm is by differentiating the deployment priority of application load feature and SLA, in the case of guaranteeing that the energy consumption of whole data center is minimum, VM application load is deployed on ready-portioned PM.

Above-mentioned dispositions method, training data comes from the real load of IaaS cloud platform, and is at least 1000 groups.

Further, in described step S3, adopt the weights and the threshold value that after training, obtain to carry out the initialization of sorter model, and then obtain the result of classification;

The precision of this result do not reach IaaS cloud platform provide service SLA require time, repeat the training of sorter model, until obtain satisfactory weights and threshold value.

The invention has the beneficial effects as follows:

(1) set up a sorter model, adopt trial of the present invention to dispose energy consumption algorithm TAEA, differentiate the deployment priority of application load feature and two factors of SLA, pay the utmost attention to which factor VM application load is deployed to ready-portioned PM above, thereby make the energy consumption of whole data center reach minimum;

(2) VM migration can respond quickly according to the variation of application load, makes energy consumption remain on optimum state;

(3) have the ability of processing the multiple SLA application of multi-user, support different types of application and mixed load, this programme only needs that it is carried out to load characteristic extraction and then classifies according to the algorithm of grader;

(4) adopt a large amount of real load data to carry out sorter model and train, the effect of carrying out optimum allocation to reach Simulation with I aaS cloud platform in practice.

Accompanying drawing explanation

Fig. 1 is the application load grader structural representation of the resource load classifier design method based on IaaS cloud platform of the present invention;

Fig. 2 is application load feature extraction and the pattern recognition flow chart of the resource load classifier design method based on IaaS cloud platform of the present invention;

Fig. 3 is the disaggregated model figure of the application load grader of the resource load classifier design method based on IaaS cloud platform of the present invention;

Fig. 4 is that the BP neural net of the resource load classifier design method based on IaaS cloud platform of the present invention builds schematic diagram;

Fig. 5 is that the application load grader TAEA algorithm of the resource load classifier design method based on IaaS cloud platform of the present invention is realized schematic diagram.

Embodiment

The following description and drawings illustrate specific embodiment of the invention scheme fully, to enable those skilled in the art to put into practice them.Embodiment only represents possible variation.

Application load is embodying user to be bought after VM, within a certain period of time, and the situations such as cpu busy percentage, internal memory use.And resource distribute refer to cloud computing provider client buy VM application deployment on the PM resource of what type.

Under the prerequisite of considering energy consumption, carry out the deployment of VMs, it is a pair of key factor that is mutually related that application load and resource are distributed.The present invention, by user's application load is analyzed, provides a solution, and the VM application of arbitrary application load or arbitrary class application load can be deployed on most suitable PM to energy consumption perception.

As shown in Figure 1, a kind of resource load classifier design method based on IaaS cloud platform of the present invention, use the statistics scientific principles such as IRQ to extract load Sampling characters, create BP neural network model based on load characteristic, utilize real load data to carry out the training of this model, after convergence, obtain weights and the threshold value of each layer of cloud platform, and then draw the optimum distribution scheme of resource, comprise particularly following step S1-S4.

Wherein, step S1 is specially analytical applications load characteristic and SLA level characteristics, and carries out extraction and the pattern recognition of load Sampling characters, and its result comprises characteristic and the classification mode that preliminary treatment obtains.

Application load and SLA are analyzed, are considered following two problems:

(1) carry out the signature analysis of resource classification deployment according to VM application load;

(2) according to the SLA grade of service signature analysis that carries out the deployment of VM application class of signing with user; Therefore for application load feature extraction is prepared.

Wherein prepare to comprise following operation:

(1) determine the frequency of load sampling and the signature identification of time window and SLA rank;

(2) determine the attribute and the form that extract feature.

In the present embodiment, adopt the statistical theories such as interquartile-range IQR (IRQ) to carry out the extraction of load Sampling characters, guarantee to embody the feature of original load sampling curve.The extraction of application load feature is the effective means that embodies the Changing Pattern of VM application load within cycle a period of time, and the classification mode that the feature that pattern recognition is extraction obtains after mating with the reference model of expectation.

Step S2 wherein extracts and creates a BP neural net based on load characteristic, input vector X using the described characteristic that obtains in step S1 as this BP neural net, output vector Y using the described classification mode that obtains in step S1 as this BP neural net, the structure of BP neural net is drawn by feature extraction and the pattern matching of system input and output.

The mapping relations between X and Y can be learnt and store to this BP neural net, the different characteristic value of taking regulated and connected weights by training, and constantly adjust weights and the threshold value of network by reflections propagate, makes the error sum of squares minimum of network.

As shown in Figure 2, the process of application load feature extraction and pattern recognition is:

First, dispose after VM application, within a time cycle, obtain load data, this load data uses mathematical method to carry out feature extraction after preliminary treatment, extract load characteristic data, the characteristic extracting, as the load characteristic pattern of this VM application, is the input vector X of BP neural net.

Then, this load characteristic pattern and reference model are compared, obtain best match pattern, this match pattern is the classification mode of this application load, and this classification mode is also the output vector Y of BP neural net simultaneously.

Particularly, before processing, can extract load sample, obtain load sample set, use IRQ statistical method for the load sample set after normalized, carry out the feature extraction of BP neural net.The present embodiment carries out illustrating of four characteristic parameters, calculate respectively load average (mean), / 4th points of positions (Q1) of load, the median (median) of load and 3/4ths points of positions (Q3) are as four features of load data, and then be combined into the input vector X={mean of this BP neural net, Q1, median, Q3}.

Then, carry out pattern matching.In training set, each sample will have a desired output, the mapping relations of setting up, the pattern matching also referred to as ANN of mating of this input and output.

For example: X～Y, and

$Y=\left\{\begin{array}{cccc}1& 0& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right.$ Represent to be output as the classification mode of four classes, like this, after recognition mode is processed, obtain as X'～Y'=(mean, Q1, median, Q3)～(1,0,0,0) input-output adapt ation model results.Like this, through application load feature extraction and initialization, as the input basis of resource load sorter model.

Wherein, step S3 creates sorter model based on described BP neural net, and it is trained and draws each layer of optimum weights and threshold value of sorter model.

As shown in Figure 3, the resource load classification model construction based on BP neural net comprises the structure of network, training and network class three parts of network.The structure of network can tentatively be determined the structure of BP neural net according to the corresponding relation of system input and output and feature extraction and pattern matching.

Determine the dimension of input signal according to load data feature, take resource load number of categories to be sorted as output node number, this BP neural net is set as single hidden layer, thereby has determined the structure of sorter model network.

As shown in Figure 4, in the present embodiment, the structure of definite BP neural net is 4-6-4, that is: 4 nodes of input layer; 6 nodes of hidden layer; 4 nodes of output layer.

After sorter model has created, need to train grader, the training data that the present embodiment adopts is all from the real load of IaaS cloud platform, be chosen at random the load data of 10 days between in March, 2011 to April, carry out 288 sampling at 5 minutes intervals of 24 hours every days, the sample set (as cpu busy percentage) that obtains training, its training data group number is much larger than 1000 groups.

The data acquisition system of training will guarantee enough quantity, the number of times of training is abundant, can have certain breadth and depth, reaches the effect that convergence has, and data are truly more conducive to IaaS cloud platform and move later scheduling of resource, obtain classification results more accurately.

Final disaggregated model training finishes, and can obtain each layer (as shown in Figure 5, comprising core exchange layer, Guinier-Preston zone network and access layer network) optimum weights and threshold value and save, and carries out load data classification.

Step S4 wherein, for using TAEA algorithm, dispatches the later application load of classifying, and this application load is deployed on most suitable server, realizes the optimum allocation of virtual machine.

After obtaining optimum weights and threshold value, need first sorter model to be carried out to initialization, it is mainly that weights and threshold value to sorter model carried out initialization, in the present embodiment, adopt random function to produce, the Initial value choice of weights and threshold value is different, can there is impact to convergence rate, but classification results is not affected.

Be applied load characteristic of correspondence classification of disaggregated model, is effectively deployed to PMs corresponding application and goes up, and reaches the object of energy efficient.Here the resource allocation algorithm of using a grader triggers scheduling of resource, realizes VMs application deployment on IaaS cloud platform to the optimization on PMs, i.e. the optimization of resource utilization,

Reach IaaS cloud platform and save the target of energy consumption.

Such grader resource allocation algorithm is called attempts disposing energy consumption algorithm TAEA (Trying allocating energy algorithm).

TAEA can differentiate the deployment priority of application load feature and two factors of SLA, pays the utmost attention to which factor VM application load is deployed to ready-portioned PMs above, makes the energy consumption of whole data center minimum.

Particularly, as shown in Figure 5, its formula is

$E=\mathrm{getPower}(\frac{\text{avg}\left(U_{\mathrm{workload}}\right)\·{\mathrm{TotalMips}}_{\mathrm{VM}}}{{\mathrm{TotalMips}}_{\mathrm{PM}}}\×100\%)*T---\left(1\right)$

In formula (1), E is the upper energy consumption in time T of VM application deployment to PM, and getPower () is the linear energy consumption calculation algorithm from Cloudsim, avg (U _workload) be the average of VM application load in time T, TotalMips _vMthe TotalMips of VM type, TotalMips _pMbe the interior check figure (core numbers) of dominant frequency X of corresponding PM, T is the time of investigating VM application load.

Experimental data derives from ten days real load data that PlanetLab project provides, through each 200 experiment porchs of building of four class servers as shown in table 1, get the T=86400 result that second, (1 day) tested gained, the percentage of the cpu frequency of the line display different server in table 1, the corresponding energy consumption (when per kilowatt) consuming when server CPU percentage is shown in the list in table.

Table 1 derives from the server energy consumption data that SPEC tissue is announced

Use after this programme and do not use the data of the aspect such as energy consumption, resource optimization of cloud platform of this programme as shown in table 2 below, data in contrast table 2 can be found out, use IaaS cloud platform after TAEA algorithm of the present invention Energy Intensity Reduction more than 40 percent, its resource optimization rate to application load has promoted more than 20 percent.

Table 2 experimental data of the present invention

Above-described embodiment is only for the invention example is clearly described, and the not restriction to the invention embodiment.For those of ordinary skill in the field, can also make other changes in different forms on the basis of the above description.Here without also giving exhaustive to all execution modes.All any apparent variations of being extended out within the spirit and principles in the present invention or variation are still among the protection range in the invention claim.

Claims (10)

1. the resource optimization dispositions method based on IaaS cloud platform, is characterized in that, comprises the following steps:

Step S1: analytical applications load characteristic and SLA level characteristics, carry out extraction and the pattern recognition of load Sampling characters, its result comprises characteristic and the classification mode that preliminary treatment obtains;

Step S2: extract and create a BP neural net based on load characteristic, its input vector X is described characteristic, and its output vector Y is described classification mode, is drawn the structure of this BP neural net by the feature extraction of system input and output and pattern matching;

Step S3: create sorter model based on described BP neural net, and it is trained and draws each layer of optimum weights and threshold value of sorter model;

Step S4: use TAEA algorithm, the later application load of classifying is dispatched, this application load is deployed on most suitable server, realize the optimum allocation of virtual machine.

2. dispositions method according to claim 1, is characterized in that, in described step S1, adopts the statistical theories such as IRQ interquartile-range IQR to carry out the extraction of load Sampling characters, to guarantee to embody the feature of original load sampling curve.

3. dispositions method according to claim 1, is characterized in that, described load Sampling characters at least comprises cpu busy percentage and the memory usage of a period of time with interior application load, and described load Sampling characters is carried out to following operation:

(1) determine the signature identification of frequency, time window and the SLA rank of load sampling;

(2) determine the attribute and the form that extract feature.

4. according to the arbitrary described dispositions method of claims 1 to 3, it is characterized in that, described load Sampling characters is formulated and initialization by the keeper of cloud platform.

5. dispositions method according to claim 1, it is characterized in that, described BP neural net is single hidden layer, it is take resource load number of categories to be sorted as output node number, and the resource load grader modeling based on BP neural net comprises structure, the training of network and three parts of the classification of network of network.

6. dispositions method according to claim 5, is characterized in that, the convergence rate of described sorter model is more than or equal to IaaS cloud platform provides the SLA requirement of service.

7. dispositions method according to claim 1, is characterized in that, triggers scheduling of resource by the resource allocation algorithm of described sorter model, according to the classification of application load characteristic of correspondence, corresponding application load is deployed on PM effectively.

8. dispositions method according to claim 1, it is characterized in that, in described step S4, described TAEA algorithm is by differentiating the deployment priority of application load feature and SLA, in the case of guaranteeing that the energy consumption of whole data center is minimum, VM application load is deployed on ready-portioned PM.

9. dispositions method according to claim 1, is characterized in that, training data comes from the real load of IaaS cloud platform, and is at least 1000 groups.

10. dispositions method according to claim 9, is characterized in that, in described step S3, adopts the weights and the threshold value that after training, obtain to carry out the initialization of sorter model, and then obtains the result of classification;

The precision of this result do not reach IaaS cloud platform provide service SLA require time, repeat the training of sorter model, until obtain satisfactory weights and threshold value.

Images