CN102279771A - Method and system for adaptively allocating resources as required in virtualization environment - Google Patents

Abstract

Translated fromChinese

一种虚拟化环境中自适应按需资源分配的系统，它包括动态感知请求分发模块、1-PM:N-VM模块和数据中心全局管理模块。1-PM:N-VM模块根据实时收集的用户体验进行一台物理机上资源的分配；动态感知请求分发模块根据监测的应用请求负载信息和虚拟机容量信息把负载分发到合适的虚拟机上对请求进行响应；数据中心全局管理模块根据收集的物理机资源负载信息决定是否需要虚拟机在物理机之间进行迁移以重新进行放置，在物理机过量或不足时是否向空闲资源池释放或申请新的物理机以退出或加入应用的服务。一种虚拟化环境中自适应按需资源分配的方法，它包括自适应的虚拟机动态容量感知请求分发策略、1-PM:N-VM资源分配策略和虚拟机迁移策略。本发明在计算机技术领域里有应用前景。

A system for self-adaptive on-demand resource allocation in a virtualized environment, which includes a dynamic perception request distribution module, a 1-PM:N-VM module and a data center global management module. 1-PM: The N-VM module allocates resources on a physical machine according to the user experience collected in real time; the dynamic perception request distribution module distributes the load to the appropriate virtual machine according to the monitored application request load information and virtual machine capacity information. Respond to the request; the data center global management module decides based on the collected physical machine resource load information whether virtual machines need to be migrated between physical machines for re-placement, and whether to release or apply for new ones to the idle resource pool when the physical machines are excessive or insufficient. The physical machine to exit or join the application's service. A method for adaptive on-demand resource allocation in a virtualization environment, which includes an adaptive virtual machine dynamic capacity-aware request distribution strategy, a 1-PM:N-VM resource allocation strategy and a virtual machine migration strategy. The invention has application prospects in the field of computer technology.

Description

The self-adaptation method and system of resources allocation as required in a kind of virtualized environment

(1) technical field

The present invention relates to resources allocation, relate in particular in a kind of virtualized environment the self-adaptation method and system of resources allocation as required, belong to field of computer technology.

(2) background technology

Along with the develop rapidly that Internet uses, increasing service provider gives data center service outsourcing, makes the scale of data center become increasing, become increasingly complex.Intel Virtualization Technology makes it bring into play more and more important effect in the heart in making up large-scale data owing to have following characteristics: the resource utilization that 1, promotes architecture.Can move many virtual machines (VM) simultaneously on the physical machine (PM), monitor of virtual machine (VMM) provides more fine-grained resource allocation mechanism support, for example the credit scheduling strategy of xen (a kind of virtualization product of increasing income) acquiescence provides the mechanism support of cpu resource dynamic assignment, balloon driver provides the support of memory resource dynamic distribution mechanism, and fine-grained resources allocation approaches optimized distribution; 2, better safety, isolation.A virtual machine is subjected to virus, security threat or application wherein and causes system crash can not influence the operation of other virtual machines; 3, has better high availability.When the physical machine accident that virtual machine the moved machine of delaying, can use virtual machine (vm) migration mechanism that whole virtual machine fast transferring is continued operation to other physical machine; 4, can better reduce power consumption, refrigeration and machine room space.Utilize the Server Consolidation technology, reduce the quantity of data center's physical server and related hardware, from reduce machine room space, energy consumption and refrigeration demand, benefit; 5, improved IT employee's work efficiency.Reaching the standard grade of a new server needs complicated flow process, and newly-built virtual machine is the thing of a few minutes, has simplified the supply of server greatly.

Resource management in the data center divides three levels: how to divide the physical machine resource between a plurality of virtual machines on (1) same physical machine; (2) the placement problem of virtual machine on physical machine; (3) externally provide more than one of the server of service, how ask distribution between the virtual machine of same services providing.

In present resource management system,, exist following several problem for the distribution and the management of resource:

1) these three separate considerations of level do not have consideration dependence each other.For example, request distribution policy in the level (3) design unreasonable can not perceive the dynamic change of resources of virtual machine, the phenomenon that " futile effort " in the level (1) increases or reduce resources of virtual machine can occur.In level (1) accurately by being required to be the virtual machine Resources allocation, can be so that virtual machine be placed on the physical machine of less amount in the level (2).

2) the expectation resource utilization that all is based on definition at the Tivoli of DRS, the IBM of HP-UX Workload Manager, the VMware of level (1) Hewlett-Packard Corporation at present excites resource to redistribute adjustment.At the P.Padala of academia, X.Zhu etc. also are based on the adjustment that similar thinking is carried out resource.Exist following shortcomings based on resource utilization adjustment: (a) along with the Internet application and development, use and to become more and more " huge ", to become increasingly complex, what suitablely become more and more difficult in the resource utilization definition of expectation.If definition is too high, can influence the performance of application.If definition is low excessively, cause wasting resource.(b) under virtualized environment, it is difficult more that resource utilization sampling accurately becomes, because need to filter the expense that monitor of virtual machine brings.(c) when a plurality of virtual machines run on the same physical machine, predict that with the current resource utilization that records real resource requirement accuracy is very poor.

3) traditional no longer suitable under virtualized environment at the request distribution policy under the physical machine environment.Under the physical machine environment, remove the non-administrator and carry out scale up or scale down by hand, otherwise it is constant using shared server resource capacity in operational process, and under virtualized environment, providing dynamically more fine granularity resource allocation mechanism support owing to VMM, dynamic change can take place in the resource capacity of virtual machine in using operational process.

4) under the conventional situation for reduce power consumption and the refrigeration cost, obsolete physical machine is shut down to reach purpose of energy saving.But web service has very big suddenly under the actual conditions, and it is too big to add the expense that new physical machine participates on working time when this situation takes place.According to statistics, if system can not respond the satisfaction that exceeds the user to user's request on time, the service provider not only is faced with " lifting " user to serving the risk of dejected degree, What is more, and the user may turn to rival's service, influence product and company's reputation of company, bring very big economic loss.

In order to solve the problem of above existence, we have designed cover self-adaptation resource allocation system as required.At problem 2), carry out the foundation that resource is redistributed between a plurality of virtual machines on the physical machine at level (1) according to user's the Expected Response time, the more direct reality that more meets of this mode.At problem 3), designed the dynamic change of the dynamic sensing module energy of a resources of virtual machine capacity real-time perception virtual machine capacity at level (2).At problem 4), safeguarded an idling-resource pond, when physical machine is not used adding idling-resource pond, not to allow it shut down immediately, keep a period of time but allow it be in dormant state.At problem 1) our designed system considers interdepending between these three levels.

(3) summary of the invention

The purpose of this invention is to provide in a kind of virtualized environment the self-adaptation method and system of resources allocation as required, satisfy user experience, more effective use resource with more effective.

The present invention is achieved by the following technical solutions:

1, self-adaptation resource distributor system as required in a kind of virtualized environment of the present invention, it comprises distribution according to need module between a plurality of virtual machines on dynamic perception request distribution module, the physical machine (being called for short the 1-PM:N-VM module) and global administration of data center module, as shown in Figure 1.Relation between them is: 1-PM:N-VMmodule 101 is carried out the distribution of resource on the physical machine according to the user experience of real-time collecting; Dynamically perceptionrequest distribution module 102 responds load distribution to suitable virtual machine according to the application request load information and the virtual machine capacity information of monitoring to request; Whether whether global administration ofdata center module 103 determines the needs virtual machine to move between physical machine placing again according to the physical machine resource load information of collecting, discharge or apply for that new physical machine is to withdraw from or to add the service of application to the idling-resource pond when physical machine is excessive or not enough.

Described dynamic perception request distribution module comprises: application load monitoring module, application load prediction module, the dynamic sensing module of virtual machine capacity and request distribution module, as shown in Figure 2.Relation between them is: applicationload monitoring module 201 monitoring request load informations pass to applicationload prediction module 202 to the result; Applicationload prediction module 202 is used long-term forecasting to add the short-term modification method charge capacity is predicted; The capacity information ofdynamic sensing module 203 each virtual machines of real-time collecting of virtual machine capacity;Request distribution module 204 selects suitable virtual machine processing request that it is responded according to the result of load estimation and the capacity information of each virtual machine.

This application load monitoring module is: each class is used all needs a load distribution device, and the load monitoring module is moved thereon.All requests at first arrive corresponding load distribution device, and application load monitoring module herein can be known overall load information easily.This module is a program element.

This application load prediction module is: the method for using long-term forecasting to add the short-term correction is predicted load.The load condition of predicting i+1 stage according to the actual measurement load and the predicated error in i stage in i stage of application load monitoring module.This module is a program element.

The dynamic sensing module of this virtual machine capacity is: monitoring in real time moves the volume change of each virtual machine on it in the domain0 control desk virtual machine in physical machine.This module is a program element.

This request distribution module is: according to designed strategy request is forwarded in the corresponding virtual machine and handles.This module is a program element.

Described distribution according to need module is that the 1-PM:N-VM module comprises: request response time obtains module, resource requirement control module and resource ruling control module, as shown in Figure 3.Relation between them is: requestresponse time module 301 obtains to move in each virtual machine the request response time of using in real time; Resourcerequirement control module 302 determines the stock number that need increase or reduce according to the actual response time and the defined Expected Response temporal differences that obtain in real time; Each the virtual demand that resourceruling control module 303 is provided according to the resource requirement control module and the restrictive condition of total resources are made the decision of final resources allocation.

This request response time obtains module: stab time of arrival and the dispose difference of timestamp of response obtains the actual treatment time of each request at server end according to request.This module is a program element.

This resource requirement control module is: according to the actual treatment time of request and the resource requirement of Expected Response time decision virtual machine.This module is a program element.

This resource ruling control module is: according to the resource requirement of each virtual machine and the final resource allocation result of limit decision of total resources.When resource contention not taking place, by being required to be each virtual machine Resources allocation.When competition takes place when, service difference is provided, preferentially guarantee the resource requirement of high-priority applications.This module is a program element.

Global administration of described data center module comprises: physical machine monitoring resource module, virtual machine (vm) migration administration module and idling-resource pond administration module.As shown in Figure 4.Relation between them is: all kinds of resource operating positions of physical machinemonitoring resource module 401 monitors physical machines; The information that virtual machine (vm)migration administration module 402 is collected according to physical machine monitoring resource module determines whether the migration of needs generation virtual machine between each physical machine reapposes, and selects which virtual machine to move; Result after idling-resourcepond administration module 403 moves according to the virtual machine (vm) migration administration module determines whether needs are to free pool application or release physical machine.

This physical machine monitoring resource module is: all kinds of resource operating positions of monitors physical machine.This module is a program element.

This virtual machine (vm) migration administration module is: when the physical machine monitoring module monitors the resource utilization of physical machine when too high, need take some virtual machines of operation on it certain strategy to move on the lighter physical nodes of other load resource utilization rates.The resource utilization that monitors most of physical machine when the physical machine monitoring module is crossed when low, need take the low virtual machine of operation physically of some load certain strategy to move on other physical nodes, go out more physical machine with the free time and be discharged into the purpose that the idling-resource pond reaches the reduction power consumption.This module is a program element.

This idling-resource pond administration module is: be responsible for safeguarding the idling-resource pond, when the current physical machine quantity of moving " excessive " or " deficiency ", idling-resource pond administration module is responsible for to application interpolation of idling-resource pond or release physical machine.This module is a program element.

2, the self-adaptation method of resources allocation as required in a kind of virtualized environment of the present invention, comprising three concrete strategies:

1) adaptive virtual machine dynamic capacity perception request distribution policy, its concrete steps are as follows:

Step 1: monitor application load in real time;

Step 2: use the method for long-term forecasting and short-term correction, the application load of predicting i+1 stage according to the actual measurement load and the predicated error in i stage;

Step 3: the resource capacity information that obtains virtual machine in real time;

Step 4: charge capacity information and the load threshold of presetting according tostep 1 prediction compare, if less than load threshold, when application load is very low, should be discharged into the idling-resource pond to reach the purpose that reduces power consumption so that vacate more physical machine as far as possible with load centralization; If be higher than load threshold, should be as far as possible with load balancing, the virtual machine of avoiding having " hurries extremely ", the generation of the virtual machine that has " not busy dead " phenomenon.

2) 1-PM:N-VM resource allocation policy

Designed on a kind of physical machine method (being called for short the 1-PM:N-VM resource allocation policy) of carrying out resources allocation between many virtual machines according to the user experience of using (referring to the response time) at this.Its concrete steps are as follows:

Step 1: obtain to operate in the request response time of using on each virtual machine in real time;

Step 2: be the response time scope of an expectation of every class application definition.Statistics response time in each cycle falls into Expected Response time range left side, just in time falls in the Expected Response time range, falls into the request number on Expected Response time range right side.The user satisfaction of definition expectation, the user satisfaction of expectation represents that with the request number and the total ratio of number of asking that fall in the Expected Response time range high more effect is good more.When the request number that falls into Expected Response time range left side and total ratio of number of asking exceed the resource abundance that certain threshold value means to be provided for application, should reduce resource.When the request number that falls into Expected Response time range right side and total ratio of number of asking exceed the phenomenon that certain threshold value means the supply of generation inadequate resource, should increase resource;

Step 3: according to resource difference and the function model decision needs increase of response time difference or the stock number that reduces of design;

Step 4: according to the Optimization Theory Model of design, according to the resource requirement of each virtual machine and the final resources allocation of restriction ruling of total resources.

3) virtual machine (vm) migration strategy, its concrete steps are as follows:

Step 1: be the bound threshold value of a resource utilization of every physical machine definition

Step 2: do not have the resource utilization of physical machine to be higher than the upper limit again this moment when the physical machine resource utilization is lower than lower limit, and exist the virtual machine (can receive and mean that other physical machine sink virtual machines can not cause the resource utilization of physical machine to exceed the upper limit) that physical machine can receive on it operation and indicating physical machine " excessive " supply, the virtual machine (vm) migration of operation on it on the physical machine that can receive, is reduced the quantity of physical machine.When the physical machine resource utilization is higher than the upper limit, but exist physical machine can receive the virtual machine of operation on it, need on the physical machine that can receive, continue operation to virtual machine (vm) migration.When the physical machine resource utilization is higher than upper limit threshold and has kept a period of time, the virtual machine that does not have to receive operation it on this moment continues the physical machine of operation, and the physical machine adding that just should absorb from the idling-resource pond in dormancy moves.

3, the invention has the advantages that:

1) having designed one takes all factors into consideration resources allocation between a plurality of virtual machines on the physical machine, provides request distribution, virtual machine between a plurality of virtual machines of equal service to place the adaptive resource Governance framework of three levels.Designed system has adaptivity, has reduced system manager's maintenance cost.

2) design the strategy that carries out resources allocation according to the user experience of reality, carried out the method for resources allocation adjustment with traditional foundation expectation resource utilization and compare, not only directly had more realistic meaning.

3) be different from traditional request distribution policy, the dynamic sensing module of designed virtual machine capacity can real-time perception arrive the variation of virtual machine capacity, and then can better utilize resource.

" sudden change " of load can better be tackled in the idling-resource pond of 4) being safeguarded.Because the physical machine that is discharged in the idling-resource pond is not to shut down immediately, keeps a period of time but be in dormant state, reduced the start-up time that load changing dynamically adds needs.

(4) description of drawings

Fig. 1 self-adaptation of the present invention is resources allocation general frame figure as required

The dynamic perception request of Fig. 2 the present invention distribution module structural representation

Fig. 3 1-PM:N-VM resource distribution module of the present invention structural representation

Global administration of Fig. 4 data center of the present invention modular structure synoptic diagram

The dynamic perception request of Fig. 5 the present invention distribution policy process flow diagram

Figure 61-PM:N-VM resource allocation policy process flow diagram

Fig. 7 virtual machine (vm) migration strategic process figure

Symbol description is as follows among the figure:

101 1-PM:N-VM modules; 102 dynamic perception request distribution modules; 103 global administration of data center modules; 201 application load monitoring modules; 202 application load prediction module; The dynamic sensing module of 203 virtual machine capacity; 204 request distribution modules; 301 request response time obtain module; 302 resource requirement control modules; 303 resource ruling control modules; 401 physical machine monitoring resource modules; 402 virtual machine (vm) migration administration modules; 403 idling-resource pond administration modules.

PM Physical Machine, physical machine; VM Virtual Machine, virtual machine; 1-PM:N-VM represents the resources allocation between a plurality of virtual machines on the physical machine; Tij i total request number that is applied in j measurement period; Nij i is applied in the request number that j measurement period actual response time falls into the Expected Response time; Lij i is applied in the request number that j measurement period actual response time falls into Expected Response time left side; Uij i is applied in the request number that j measurement period actual response time falls into Expected Response time right side.

(5) embodiment

For making technical solution of the present invention and advantage clearer,, the technical scheme in the embodiment of the invention is carried out clear, complete description below in conjunction with the embodiment of the invention.Need to prove that in accompanying drawing or instructions, components identical is all used identical Reference numeral.

Self-adaptation resource distributor system as required in a kind of virtualized environment of the present invention, it comprises the distribution according to need module between a plurality of virtual machines on the physical machine (being called for short the 1-PM:N-VM module) 101, dynamic perceptionrequest distribution module 102, global administration ofdata center module 103, as shown in Figure 1.1-PM:N-VMmodule 101 is carried out the distribution of resource on the physical machine according to the user experience of real-time collecting; Dynamically perceptionrequest distribution module 102 responds load distribution to suitable virtual machine according to the application request load information and the virtual machine capacity information of monitoring to request; Whether whether global administration ofdata center module 103 determines the needs virtual machine to move between physical machine placing again according to the physical machine resource load information of collecting, discharge or apply for that new physical machine is to withdraw from or to add the service of application to the idling-resource pond when physical machine is excessive or not enough.

Described dynamic perceptionrequest distribution module 102 comprises: applicationload monitoring module 201, applicationload prediction module 202, thedynamic sensing module 203 of virtual machine capacity andrequest distribution module 204, as shown in Figure 2.Applicationload monitoring module 201 monitoring request load informations pass to applicationload prediction module 202 to the result; Applicationload prediction module 202 is used long-term forecasting to add the short-term modification method charge capacity is predicted; The capacity information ofdynamic sensing module 203 each virtual machines of real-time collecting of virtual machine capacity;Request distribution module 204 selects suitable virtual machine processing request that it is responded according to the result of load estimation and the capacity information of each virtual machine.

This application load monitoring module 201: each class is used all needs a load distribution device, and the load monitoring module is moved thereon.All requests at first arrive corresponding load distribution device, and application load monitoring module herein can be known overall load information easily, each charge capacity information of using of real-time collecting.This module is a program element.

This application load prediction module 202: at the load of web service, load has certain rules on long terms, but has certain suddenly in a short time again, takes long-term forecasting to add the method for short-term correction to the prediction of load.The load condition of predicting i+1 stage according to the actual measurement load and the predicated error in i stage in i stage of application load monitoring module.This module is a program element.

Thedynamic sensing module 203 of this virtual machine capacity: monitoring in real time moves the volume change of each virtual machine on it in the domain0 control desk virtual machine in physical machine, utilizes xenstore and uses the xentop order can obtain the resource capacity information of each virtual machine in real time.This module is a program element.

This asks distribution module 204: each the virtual machine capacity information according to the charge capacity and the dynamic sensing module of virtual machine capacity of the prediction of applied forcasting module are collected is forwarded to request in the corresponding virtual machine and handles according to designed strategy (seeing the request distribution policy in the real-time example one).This module is a program element.

Described distribution according to need module is that 1-PM:N-VM module 101 comprises: request response time obtainsmodule 301, resourcerequirement control module 302 and resource rulingcontrol module 303, as shown in Figure 3.Requestresponse time module 301 obtains to move in each virtual machine the request response time of using in real time; Resourcerequirement control module 302 determines the stock number that need increase or reduce according to the actual response time and the defined Expected Response temporal differences that obtain in real time; Each the virtual demand that resource rulingcontrol module 303 is provided according to the resource requirement control module and the restrictive condition of total resources are made the decision of final resources allocation.

This request response time obtains module 301: stab time of arrival and the dispose difference of timestamp of response obtains the actual treatment time of each request at server end according to request.The response time of an expectation of every class application definition.All requests all can be transmitted by the virtual bridge that is arranged in domain0, use the tshark instrument from data link layer request and respond packet to be obtained, and use four-tuple＜source IP, purpose IP, source port, destination interface〉connection of sign, analyze and obtain the response time.This module is a program element.

This resource requirement control module 302: according to the actual treatment time of request and the resource requirement of Expected Response time decision virtual machine.Stock number and the relation between the response time are very complicated, the relation between being difficult to express with a mathematical model.People such as the Wang Xiaorui of University of Tennessee confirm difference between the stock number and the difference of response time relation is easier represents with mathematical model, and the relational model of design resource difference of the present invention and response time difference draws with this needs the stock number (seeing embodiment two) that increases or reduce.This module is a program element.

This resource ruling control module 303: according to the resource requirement of each virtual machine and the final resource allocation result of limit decision of total resources.When resource contention not taking place, by being required to be each virtual machine Resources allocation.When competition takes place when, service difference is provided, preferentially guarantee the resource requirement of high-priority applications.This module is a program element.

Global administration of describeddata center module 103 comprises: physical machinemonitoring resource module 401, virtual machine (vm)migration administration module 402 and idling-resource pond administration module 403.All kinds of resource operating positions of physical machinemonitoring resource module 401 monitors physical machines; The information that virtual machine (vm)migration administration module 402 is collected according to physical machine monitoring resource module determines whether the migration of needs generation virtual machine between each physical machine reapposes, and selects which virtual machine to move; Result after idling-resourcepond administration module 403 moves according to virtual machine (vm)migration administration module 402 determines whether needs are to free pool application or release physical machine.

This physical machine monitoring resource module 401: all kinds of resource operating positions of monitors physical machine.The performance collection kit sysstat that use is increased income carries out real-time collecting to all kinds of resource operating positions of physical machine.This module is a program element.

This virtual machine (vm) migration administration module 402:, need take some virtual machines of operation on it certain strategy migration (seeing embodiment three) on the lighter physical nodes of other load resource utilization rates when the physical machine monitoring module monitors the resource utilization of physical machine when too high.The resource utilization that monitors most of physical machine when the physical machine monitoring module is crossed when low, need take the low virtual machine of operation physically of some load certain strategy (seeing embodiment three) to move on other physical nodes, go out more physical machine with the free time and be discharged into the purpose that the idling-resource pond reaches the reduction power consumption.This module is a program element.

This idling-resource pond administration module 403: be responsible for safeguarding the idling-resource pond, when the current physical machine quantity of moving " excessive " or " deficiency ", idling-resource pond administration module is responsible for to application interpolation of idling-resource pond or release physical machine.This module is a program element.

2, the present invention a kind of in virtualized environment the method for self-adaptation distribution according to need resource, comprising three concrete strategy: embodiment one, adaptive virtual machine dynamic capacity perception request distribution policy; Embodiment two, the 1-PM:N-VM resource allocation policy; Embodiment three, the virtual machine (vm) migration strategy.

1) embodiment one: adaptive virtual machine dynamic capacity perception request distribution policy

The dynamic method flow diagram of perception request distribution policy in the virtualized environment that Fig. 5 provides for the embodiment of the invention one, in the present embodiment, dynamically the method step of perception request distribution policy is:

Step 1: the load state information of using application load monitoring modular 201 monitoring current application;

Step 2: the Forecasting Methodology of using long-term forecasting to add the short-term correction is predicted the load size in next stage according to current load state information and current predicated error

Step 3: thedynamic sensing module 203 real-time perception virtual machine volume change of virtual machine capacity, safeguard one this serve the double-linked circular list of all virtual machine capacity.Double-linked circular list sorts to virtual machine according to capacity order from big to small.

Step 4: requestdistribution module 204 is forwarded to request on the suitable servers with the virtual machine capacity that service is provided according to current load condition and serves.If what charge capacity less than certain threshold value, should be tried one's best load concentrates.Calculate by following formula 1:

W₁* VM quantity+W of having of PM₂* VM resource utilization+W₃* VM resource capacity (formula 1)

Value is carried out descending sort, handle on the virtual machine of the value of forwarding the request to maximum, so that " emptying " load on it as early as possible of the few physics function of the virtual machine number of being moved, reach as early as possible the release physical machine to the idling-resource pond to reach the purpose that reduces power consumption.If charge capacity is higher than certain threshold value, that load should be tried one's best carries out equilibrium between each virtual machine.Calculate by following formula 2:

W₁* resource utilization+W of VM place PM₂* VM resource utilization+W₃/ VM resource capacity (formula 2)

Value is sorted by ascending order, handle on the virtual machine of the value of forwarding the request to minimum, so that the equilibrium that load can be tried one's best.

2) embodiment two: the 1-PM:N-VM resource allocation policy

The process flow diagram of the 1-PM:N-VM resource allocation policy that Fig. 6 provides for the embodiment of the invention two, in the present embodiment, the method step of 1-PM:N-VM resource allocation policy is:

Step 1: use request response time to obtain the actual response time thatmodule 301 obtains each request.According to＜source IP, purpose IP, source port, destination interface〉indicate a connection, utilize the difference that request time stabs and the response time stabs to obtain the actual response time of each request, and add up each and use the handling capacity T of i in j cycle_Ij, the response time of setting expectation is

Counting response time falls in the Expected Response time range

Request count N_IjFall into the response time

L is counted in the request in left side_IjOr the response time falls into

U is counted in the request on right side_Ijd_IjBe j interior actual measurement response time in cycle of i virtual machine, R_IjIt is the resource in j cycle of i virtual machine.

Step 2: setting the satisfaction of using is

P_Ij=N_Ij/ T_Ij(formula 3)

The worst tolerances of using i is, when

The satisfaction of Shi Yingyong is met.

When

When (formula 4), fall into the request number in Expected Response time range left side and the ratio of total request number and exceed the resource abundance that preset threshold means to be provided for application, user experience is fine.But this situation also is not expect the phenomenon that takes place in the middle of reality, because application takies many more resources and need pay more cost, the service provider does not obtain because better user experience is provided than falling into more income in the Expected Response time range, the net proceeds of provider is to have reduced in fact in this case, so should reduce resource.

When

When (formula 5), fall into the request number on Expected Response time range right side and the ratio of total request number and exceed the phenomenon that preset threshold means the supply of generation inadequate resource, user experience is affected, and should increase resource for it.

Step 3: the difference between the stock number and the difference of response time relation is easier to be represented with mathematical model, and the relational model of design resource difference of the present invention and response time difference draws with this needs the stock number that increases or reduce.Approaching more from the Expected Response time range, need the stock number of increase or minimizing more little.

Δ R_Ij=λ (1-P_Ij) * Δ d_Ij(formula 6)

Δ R wherein_Ij=R_Ij-R_Ij-1, Δ d_Ij=| d_Ij-D_Ij|, λ is a coefficient.

Step 4: according to the resource requirement of each virtual machine, consider the restriction of total resources, when resource contention not taking place, distribute as required, the preferential resource requirement that guarantees high-priority applications when resource contention takes place.

3) embodiment three: the virtual machine (vm) migration strategy

The process flow diagram of the virtual machine (vm) migration strategy that Fig. 7 provides for the embodiment of the invention three, the method step of virtual machine (vm) migration strategy is in the present embodiment:

Step 1: use physical machinemonitoring resource module 401 to obtain the resource operating position of each physical machine.Bound for a resource utilization of each physical machine definition.

Step 2: judge that resource utilization is greater than the upper limit?

If greater than the upper limit, be calculated as follows:

W₁* VM amount of ram+W₂* the dirty page or leaf of VM internal memory rate+W₃The cpu load of/VM (formula 7)

The virtual machine of selective value minimum moves.This is because test analysis factually, and the expense of virtual machine (vm) migration is mainly relevant with the dirty page or leaf rate that virtual machine produces in operational process with the shared amount of ram size of virtual machine.Move the expense in the time of to reduce virtual machine (vm) migration in such a way, save time.

Can judge whether to exist the physical machine that sink virtual machine continues operation? can sink virtual machine, be meant that virtual machine (vm) migration in the past after, can not make that the resource utilization of target physical machine exceeds the ceiling restriction of utilization rate.

If there is the operation of then moving.

If do not exist, then apply for adding new physical machine and put into operation to idling-resource pond administration module.

Can, judge whether to exist the physical machine that receive all virtual machines that move on the physical machine? if less than the upper limit

If exist, then carry out migration operation.

Can virtual machine (vm) migrations all on the physical machine be finished? if can, then the physical machine of free time this moment is discharged in the idling-resource pond.

Claims (2)

Translated fromChinese

1.一种虚拟化环境中自适应按需资源分配的系统，其特征在于：它包括动态感知请求分发模块、一台物理机上多个虚拟机之间的按需分配模块即1-PM:N-VM模块和数据中心全局管理模块，1-PM:N-VM模块根据实时收集的用户体验进行一台物理机上资源的分配；动态感知请求分发模块根据监测的应用请求负载信息和虚拟机容量信息把负载分发到合适的虚拟机上对请求进行响应；数据中心全局管理模块根据收集的物理机资源负载信息决定是否需要虚拟机在物理机之间进行迁移以重新进行放置，在物理机过量或不足时是否向空闲资源池释放或申请新的物理机以退出或加入应用的服务；1. A system for self-adaptive on-demand resource allocation in a virtualized environment, characterized in that: it includes a dynamic perception request distribution module, an on-demand allocation module between a plurality of virtual machines on a physical machine, i.e. 1-PM:N -VM module and data center global management module, 1-PM:N-VM module allocates resources on a physical machine according to user experience collected in real time; dynamic perception request distribution module requests load information and virtual machine capacity information according to monitored applications Distribute the load to the appropriate virtual machine to respond to the request; the data center global management module decides whether the virtual machine needs to be migrated between physical machines for relocation according to the collected physical machine resource load information, and when the physical machine is excessive or insufficient Whether to release or apply for a new physical machine to the free resource pool to exit or join the application service;所述的动态感知请求分发模块包括：应用负载监控模块、应用负载预测模块、虚拟机容量动态感知模块和请求分发模块，应用负载监控模块监测请求负载信息，把结果传递给应用负载预测模块；应用负载预测模块使用长期预测加短期修正方法对负载量进行预测；虚拟机容量动态感知模块实时收集各个虚拟机的容量信息；请求分发模块根据负载预测的结果和各个虚拟机的容量信息选择合适的虚拟机处理请求对其进行响应；The dynamic perception request distribution module includes: an application load monitoring module, an application load prediction module, a virtual machine capacity dynamic perception module and a request distribution module, the application load monitoring module monitors request load information, and transmits the result to the application load prediction module; The load prediction module uses long-term prediction plus short-term correction method to predict the load; the virtual machine capacity dynamic sensing module collects the capacity information of each virtual machine in real time; the request distribution module selects the appropriate virtual machine according to the load prediction result and the capacity information of each virtual machine The machine processes the request and responds to it;该应用负载监控模块是：每一类应用都需要一台负载分发器，负载监控模块运行于其上；所有的请求首先到达相应的负载分发器，此处的应用负载监控模块可以很容易的获知全局负载信息，该模块是个程序元件；The application load monitoring module is: each type of application needs a load distributor, and the load monitoring module runs on it; all requests first arrive at the corresponding load distributor, and the application load monitoring module here can easily know Global load information, the module is a program element;该应用负载预测模块是：使用长期预测加短期修正的方法对负载进行预测；根据应用负载监控模块第i个阶段的实测负载和第i个阶段的预测误差预测第i+1个阶段的负载状态，该模块是个程序元件；The application load prediction module is: use long-term prediction plus short-term correction method to predict the load; predict the load status of the i+1th stage according to the measured load of the i-th stage of the application load monitoring module and the prediction error of the i-th stage , the module is a program element;该虚拟机容量动态感知模块是：在物理机中的domain0控制台虚拟机中实时监测运行其上的各个虚拟机的容量变化，该模块是个程序元件；The virtual machine capacity dynamic perception module is: in the domain0 console virtual machine in the physical machine, real-time monitoring of the capacity changes of each virtual machine running on it, the module is a program component;该请求分发模块是：根据所设计的策略把请求转发到相应的虚拟机中进行处理，该模块是个程序元件；The request distribution module is: according to the designed strategy, the request is forwarded to the corresponding virtual machine for processing, and the module is a program element;所述的按需分配模块即1-PM:N-VM模块包括：请求响应时间获得模块、资源需求控制模块和资源裁决控制模块，请求响应时间模块实时获得每个虚拟机中所运行应用的请求响应时间；资源需求控制模块依据实时获得的实际响应时间和所定义的期望响应时间的差异，决定需要增加或减少的资源量；资源裁决控制模块根据资源需求控制模块所提供的各个虚拟需求和资源总量的限制条件，做出最终资源分配的决定；The on-demand distribution module, that is, the 1-PM:N-VM module, includes: a request response time acquisition module, a resource demand control module and a resource arbitration control module, and the request response time module obtains the request of the running application in each virtual machine in real time Response time; the resource demand control module determines the amount of resources that need to be increased or decreased according to the difference between the actual response time obtained in real time and the defined expected response time; the resource arbitration control module controls the various virtual demands and resources provided by the resource demand control module Amount constraints to make final resource allocation decisions;该请求响应时间获得模块是：根据请求到达时间戳和响应处理完毕时间戳的差值获得每个请求在服务器端的实际处理时间，该模块是个程序元件；The request response time obtaining module is: obtain the actual processing time of each request on the server side according to the difference between the request arrival time stamp and the response processing completion time stamp, this module is a program element;该资源需求控制模块是：根据请求的实际处理时间和期望响应时间决定虚拟机的资源需求，该模块是个程序元件；The resource demand control module is to determine the resource demand of the virtual machine according to the actual processing time and expected response time of the request, and this module is a program element;该资源裁决控制模块是：根据各个虚拟机的资源需求和资源总量的限制决定最终的资源分配结果；当不发生资源竞争时，按需为各个虚拟机分配资源；当发生竞争时，提供服务差异，优先保证高优先级应用的资源需求；该模块是个程序元件；The resource arbitration control module is to determine the final resource allocation result according to the resource requirements of each virtual machine and the limit of the total amount of resources; when there is no resource competition, allocate resources to each virtual machine on demand; when competition occurs, provide services Differences, priority is given to ensuring the resource requirements of high-priority applications; this module is a program element;所述的数据中心全局管理模块包括：物理机资源监控模块、虚拟机迁移管理模块和空闲资源池管理模块，物理机资源监控模块监控物理机的各类资源使用情况；虚拟机迁移管理模块根据物理机资源监控模块收集的信息决定是否需要发生虚拟机在各个物理机之间的迁移进行重新放置，以及选择哪些虚拟机进行迁移；空闲资源池管理模块根据虚拟机迁移管理模块进行迁移后的结果决定是否需要向空闲池申请或释放物理机；The data center global management module includes: a physical machine resource monitoring module, a virtual machine migration management module and an idle resource pool management module, the physical machine resource monitoring module monitors the usage of various resources of the physical machine; the virtual machine migration management module The information collected by the machine resource monitoring module determines whether virtual machines need to be migrated between physical machines for relocation, and which virtual machines are selected for migration; the idle resource pool management module decides based on the migration results of the virtual machine migration management module Whether it is necessary to apply for or release physical machines from the free pool;该物理机资源监控模块是：监控物理机的各类资源使用情况，该模块是个程序元件；The physical machine resource monitoring module is to monitor the usage of various resources of the physical machine, and this module is a program element;该虚拟机迁移管理模块是：当物理机监控模块监测到物理机的资源使用率过高时，需要把其上运行的一些虚拟机采取一定的策略迁移到其它负载资源使用率较轻的物理结点上；当物理机监控模块监测到大多数物理机的资源使用率过低时，需要把某些负载低的物理上运行的虚拟机采取一定的策略迁移到其他物理节点上，以空闲出更多的物理机释放到空闲资源池达到降低功耗的目的，该模块是个程序元件；The virtual machine migration management module is: when the physical machine monitoring module detects that the resource utilization rate of the physical machine is too high, it needs to adopt a certain strategy to migrate some virtual machines running on it to other physical structures with light load resource utilization rate. point; when the physical machine monitoring module detects that the resource usage of most physical machines is too low, it is necessary to adopt a certain strategy to migrate some virtual machines with low loads to other physical nodes to free up more resources. Release more physical machines to the free resource pool to reduce power consumption. This module is a program component;该空闲资源池管理模块是：负责维护空闲资源池，当前正在运行的物理机数量“过量”或“不足”时，空闲资源池管理模块负责向空闲资源池申请添加或释放物理机，该模块是个程序元件。The idle resource pool management module is responsible for maintaining the idle resource pool. When the number of currently running physical machines is "excessive" or "insufficient", the idle resource pool management module is responsible for applying to the idle resource pool to add or release physical machines. This module is a program element.2.一种虚拟化环境中自适应按需资源分配的方法，其特征在于：它包括三个具体的策略：2. A method for adaptive on-demand resource allocation in a virtualized environment, characterized in that: it includes three specific strategies:1)自适应的虚拟机动态容量感知请求分发策略，其具体步骤如下：1) Adaptive virtual machine dynamic capacity awareness request distribution strategy, the specific steps are as follows:步骤1：实时监控应用负载；Step 1: Monitor application load in real time;步骤2：使用长期预测和短期修正的方法，根据第i个阶段的实测负载和预测误差预测第i+1个阶段的应用负载；Step 2: Using the method of long-term prediction and short-term correction, predict the application load of the i+1th stage according to the measured load and forecast error of the i-th stage;步骤3：实时获得虚拟机的资源容量信息；Step 3: obtain resource capacity information of the virtual machine in real time;步骤4：据步骤1预测的负载量信息与预设的负载阈值进行比较，若小于负载阈值，应用负载很低时，应尽量将负载集中，以便空出更多的物理机释放到空闲资源池以达到降低功耗的目的；若高于负载阈值，应尽量将负载均衡，避免出现有的虚拟机“忙死”，有的虚拟机“闲死”现象的发生；Step 4: Compare the load information predicted in step 1 with the preset load threshold. If it is less than the load threshold, when the application load is very low, try to concentrate the load so that more physical machines can be released to the idle resource pool In order to achieve the purpose of reducing power consumption; if it is higher than the load threshold, the load should be balanced as much as possible to avoid the phenomenon of "busy death" of some virtual machines and "idle death" of some virtual machines;2)1-PM:N-VM资源分配策略，其具体步骤如下：2) 1-PM:N-VM resource allocation strategy, the specific steps are as follows:步骤1：实时获得运行在每个虚拟机上应用的请求响应时间；Step 1: Obtain the request response time of the application running on each virtual machine in real time;步骤2：为每类应用定义一个期望的响应时间范围；统计在每个周期内响应时间落入期望响应时间范围左侧，正好落入期望响应时间范围内，落入期望响应时间范围右侧的请求数；定义期望的用户满意度，期望的用户满意度以落入期望响应时间范围内的请求数和总请求数的比例来表示，越高效果越好；当落入期望响应时间范围左侧的请求数和总请求数的比例超出一定的阈值意味着为应用提供的资源充足，应当减少资源；当落入期望响应时间范围右侧的请求数和总请求数的比例超出一定的阈值意味着发生资源不足供应的现象，应该增加资源；Step 2: Define an expected response time range for each type of application; statistics in each cycle that the response time falls on the left side of the expected response time range, just falls within the expected response time range, and falls on the right side of the expected response time range The number of requests; define the expected user satisfaction, the expected user satisfaction is represented by the ratio of the number of requests falling within the expected response time range to the total number of requests, the higher the effect, the better; when it falls on the left side of the expected response time range The ratio of the number of requests to the total number of requests exceeds a certain threshold, which means that the resources provided for the application are sufficient, and resources should be reduced; when the ratio of the number of requests to the total number of requests that falls on the right side of the expected response time range exceeds a certain threshold, it means In case of insufficient supply of resources, resources should be increased;步骤3：根据设计的资源差异和响应时间差异的函数模型决定需要增加或减少的资源量；Step 3: Determine the amount of resources that need to be increased or decreased according to the function model of the designed resource difference and response time difference;步骤4：依据设计的优化理论模型，根据各个虚拟机的资源需求和资源总量的限制裁决最终的资源分配；Step 4: According to the designed optimization theoretical model, determine the final resource allocation according to the resource requirements of each virtual machine and the limitation of the total amount of resources;3)虚拟机迁移策略，其具体步骤如下：3) virtual machine migration strategy, the specific steps are as follows:步骤1：为每台物理机定义一个资源使用率的上下限阈值；Step 1: Define an upper and lower limit threshold for resource usage for each physical machine;步骤2：当物理机资源使用率低于下限此时又没有物理机的资源使用率高于上限，并且存在着物理机可以接收其上运行的虚拟机，预示着物理机“过量”供应，将其上运行的虚拟机迁移到可以接收的物理机上，减少物理机的数量；当物理机资源使用率高于上限，但存在物理机可以接收其上运行的虚拟机，需要把虚拟机迁移到可以接收的物理机上继续运行；当物理机资源使用率高于上限阈值并维持了一段时间，此时没有可以接收其上运行的虚拟机继续运行的物理机，应从空闲资源池中吸收正在休眠的物理机加入运行。Step 2: When the resource utilization rate of the physical machine is lower than the lower limit, and the resource utilization rate of no physical machine is higher than the upper limit, and there are physical machines that can receive virtual machines running on it, it indicates that the physical machine is "over-provisioned". The virtual machine running on it is migrated to a physical machine that can receive it to reduce the number of physical machines; when the resource usage of the physical machine is higher than the upper limit, but there are physical machines that can receive the virtual machine running on it, the virtual machine needs to be migrated to a physical machine that can The receiving physical machine continues to run; when the resource usage of the physical machine is higher than the upper threshold and lasts for a period of time, and there is no physical machine that can receive the virtual machine running on it to continue running, the sleeping physical machine should be absorbed from the free resource pool. The machine joins the operation.

Images