CN103294546A - Multi-dimensional resource performance interference aware on-line virtual machine migration method and system - Google Patents

Abstract

The invention discloses an on-line virtual machine migration method and belongs to the technical field of computer virtualization. Particularly, the method comprises the following steps: a central control server monitors the resource consumption conditions of physical servers to optimize the resource consumption conditions to determine a migration source physical server for targets; candidate virtual machines to be migrated are selected on the migration source physical server to constitute candidate migration schemes in a one-to-one correspondence manner; overall performance interference caused by the candidate migration schemes is evaluated through migration performance interference caused by a comprehensive virtual machine migration process and same machine performance interference to a destination physical server after migration; and the candidate migration scheme with the minimum overall performance interference is selected as the final migration scheme; and the migration source physical server fulfils the on-line virtual machine migration according to the final migration scheme. The method integrates the migration performance interference and the same machine performance interference, selects the proper migration scheme by targeting the minimum interference, and effectively reduces performance interference caused by on-line virtual machine migration.

Description

The various dimensions resource performance disturbs the online moving method of virtual machine and the system of perception

Technical field

The invention belongs to computer virtualized technical field, more specifically, relate to the method and system of the online migration of virtual machine in a kind of virtualized environment.

Background technology

Current, Intel Virtualization Technology is widely used in the cluster that large-scale data center Data Center(is made up of many physical servers) in, Amazon EC2 for example is to realize efficient utilization and the flexible management of a large amount of physical server resources (various dimensions resources such as calculating, internal memory and the network bandwidth).Because the user's of data center running load has time-varying characteristics, therefore need reasonably dispatch or integrate virtual machine, namely utilize the online migrating technology of virtual machine to realize daily management behaviors such as the load balancing of data center and energy consumption saving.

Inner owing to virtual machine moves the negative performance impact of bringing between different physical servers in order to alleviate data center; the online migration scheme of existing virtual machine is mainly paid close attention to by the performance cost of migration virtual machine itself, for example moves time delay, moves factor such as stop time.Yet, these schemes ignored virtual machine in transition process with migration source physical server and target physical server on the performance that produces of virtual machine disturb (being called for short migration disturbs), and after migration with the physical server of migration purpose on the same machine performance that produces of virtual machine disturb (being called for short with the machine interference).This causes the online migration scheme of existing virtual machine can not multianalysis, quantification and alleviate the performance of being brought by virtual machine (vm) migration and disturb (being called for short the virtual machine (vm) migration performance disturbs).In addition, because the measurement of off-line type is taked in the performance interference that most of existing scheme produces virtual machine (vm) migration or only the interference of the performance on part dimension resource (as CPU, internal memory) quantizes, therefore cause the result of its quantification inadequately comprehensively accurately.This can influence migration scheme optimization decision-making (being virtual machine to be migrated and the selection of moving the target physical server), and then reduces the performance of data center.Comprehensive above-mentioned 2 points, existing virtual machine (vm) migration scheme has complicated process steps, disturb quantize inaccurate, problem such as migration performance is not good.

Summary of the invention

Defective at prior art, the invention provides the method for the online migration of virtual machine in a kind of virtualized environment, the migration performance that the virtual machine (vm) migration process of considering this method synthesis causes disturbs and migration is disturbed the same machine performance that the target physical server causes afterwards, be that target is chosen suitable migration scheme to disturb minimum, effectively reduce the online migration of virtual machine and the performance that produces is disturbed.

For realizing above-mentioned technical purpose, the invention provides the online moving method of a kind of virtual machine, may further comprise the steps:

(1) the center Control Server is monitored the resource consumption state of each physical server, is the physical server that target defines the virtual machine (vm) migration demand to optimize the resource consumption state, is designated as migration source physical server;

(2) choose candidate's virtual machine to be migrated at migration source physical server, and choose the candidate for each candidate virtual machine to be migrated and move the target physical server, the candidate moves that the target physical server need satisfy its available CPU nuclear number and memory size is enough to the virtual machine use to be migrated for the candidate, with candidate's virtual machine to be migrated and candidate move the target physical server one by one corresponding formation candidate move scheme;

(3) estimate the overall performance interference that each candidate's migration scheme causes, its process is:

More big and transit time is more long according to the network I/O bandwidth contention of migration source physical server and target physical server, then transition process disturbs more big criterion to the migration performance that migration source physical server and target physical server cause, and the migration performance that evaluate candidate migration scheme causes disturbs;

The more big then virtual machine to be migrated of network I/O resources occupation rate in conjunction with virtual machine to be migrated disturbs more big to the same machine performance that migration target physical server causes after migration, and the demand of the buffer memory of virtual machine to be migrated and memory bandwidth is more big, and the supply of buffer memory and memory bandwidth is more little on the migration target physical server, then virtual machine to be migrated disturbs more big criterion to the same machine performance that migration target physical server causes after migration, and the same machine performance that evaluate candidate migration scheme causes is disturbed;

The migration performance that comprehensive evaluation candidate's migration scheme causes disturbs and disturbs with the machine performance, obtains the overall performance interference that candidate's migration scheme causes;

(4) choosing overall performance from each candidate's migration scheme disturbs reckling as final migration scheme;

(5) migration source physical server is finished the online migration of virtual machine according to final migration scheme.

Further, in the described step (3), the virtual machine network I/O bandwidth consumption on the described network I/O bandwidth contention employing physical server and the network of physical server interrupt supply and demand ratio comprehensively to be weighed; Described transit time adopts the memory size of virtual machine to be migrated and the dirty page or leaf of the internal memory rate measurement of virtual machine to be migrated comprehensively to weigh; Described network I/O resources occupation rate adopts network I/O bandwidth or network I/O interruption times or network TCP socket number to weigh; The buffer memory of described virtual machine to be migrated and the demand of memory bandwidth adopt virtual machine cache miss rate and memory usage sum to weigh; The supply of buffer memory and memory bandwidth adopts the cache hit rate of this physical server and internal memory surplus ratio sum to weigh on the described migration target physical server.

The specific implementation of the migration performance interference that further, evaluate candidate migration scheme causes in the described step (3) is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; The candidate moves scheme＜V _l, P _dThe migration performance that causes disturbs

, ω _sBe virtual machine V to be migrated _lThe migration performance that in transition process migration source physical server is caused disturbs ω _dBe virtual machine V to be migrated _lIn transition process to migration target physical server P _dThe migration performance that causes disturbs, ω _SmThe transition process that moves scheme for all candidates disturbs maximal value, ω to the migration performance that migration source physical server causes _DmThe transition process that moves scheme for all candidates disturbs maximal value to the migration performance that migration target physical server causes;

Wherein, the migration performance that in transition process migration source or target physical server caused of virtual machine to be migrated disturbs

, α is the number of the virtual machine of migration source or the carrying of target physical server, μ _IhBe the average of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, σ _IhBe the standard deviation of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, μ _ItBe the average of the network interruption times of i virtual machine of migration source or target physical server, β _lBe the internal memory after the virtual machine normalization to be migrated, D _lBe the dirty page or leaf of the internal memory of virtual machine to be migrated rate, μ _tAverage for the network interruption times of migration source or target physical server;

The specific implementation that the same machine performance that described step (3) evaluate candidate migration scheme causes is disturbed is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; Virtual machine V to be migrated _lSame machine performance disturb

, γ _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe network bandwidth performance that causes is disturbed, ε _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe buffer memory that causes and the performance of memory bandwidth are disturbed, γ _mFor the virtual machine to be migrated in all candidate's migration schemes disturbs maximal value, ε in the network bandwidth performance that migration target physical server causes _mThe virtual machine to be migrated that moves scheme for all candidates disturbs maximal value in buffer memory and memory bandwidth performance that migration target physical server causes;

Wherein, ${\mathrm{\γ}}_{\mathrm{ld}}=\frac{{\mathrm{\μ}}_{\mathrm{lt}}}{{\mathrm{\μ}}_{\mathrm{dt}}}(\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_{\mathrm{dt}}}),$ ${\mathrm{\ϵ}}_{\mathrm{ld}}=\frac{{\mathrm{VU}}_{\mathrm{ml}}+{\mathrm{VCM}}_l}{(1-{\mathrm{PU}}_{\mathrm{md}})+(1-{\mathrm{PCM}}_d)},$

μ _LtBe virtual machine V to be migrated _lNetwork interruption times average, μ _DtBe the network interruption average of migration target physical server, α _dBe the virtual machine number of migration target physical server carrying, VU _MlBe virtual machine V to be migrated _lMemory usage, μ _ItBe the average of the network interruption times of migration target physical server i virtual machine, VU _MlBe virtual machine V to be migrated _lMemory usage, VCM _lBe virtual machine V to be migrated _lThe cache miss rate, PU _MdBe the memory usage of migration target physical server, PCM _dCache miss rate for migration target physical server;

The migration performance that the comprehensive candidate's migration scheme of described step (3) causes disturbs and disturbs with the machine performance and obtains the specific implementation that overall performance that candidate's migration scheme causes disturbs and be: overall performance disturbs T=ρ M+ (1-ρ) N, weight coefficient 0≤ρ≤1, the migration performance that M causes for candidate's migration scheme disturbs, and the same machine performance that N causes for candidate's migration scheme is disturbed.

Further, if the running load of migration source physical server and target physical server is network I/O intensity, the ρ value is greater than 0.5; If the running load of migration source physical server and target physical server is the CPU computation-intensive, the ρ value is less than 0.5; If the running load of the device of migration source and destination physics service is the mixed type load of network I/O intensity and CPU computation-intensive, ρ value 0.5.

Further, described step (1) is specially: the center Control Server is monitored the resource consumption state of each physical server, determines the physical server that resource utilization is too high or too low according to the resource consumption state; If exist many physical server resource utilizations too high, choose then that to transship severe patient be migration source physical server; If exist many physical server utilizations of resources low excessively, then choose the lowest as migration source physical server; If the too high physical server of resource utilization and the low excessively physical server of resource utilization occur simultaneously, then choose the too high physical server of resource utilization as migration source physical server; Described resource consumption state adopts CPU usage or memory usage or network I/O bandwidth consumption to weigh.

Further, the specific implementation of choosing candidate's virtual machine to be migrated at migration source physical server is: for the too high migration source physical server of resource utilization, selected part to the big virtual machine of physical server resources occupation rate as candidate's virtual machine to be migrated; For the low excessively migration source physical server of resource utilization, the virtual machine that it is all is as candidate's virtual machine to be migrated.

Another object of the present invention is to provide the online migratory system of virtual machine in a kind of virtualized environment, this system synthesis is considered the same machine performance interference that migration performance disturbs and migration causes the target physical server afterwards that the virtual machine (vm) migration process causes, be that target is chosen suitable migration scheme to disturb minimum, effectively reduce the online migration of virtual machine and the performance that produces is disturbed.

For realizing above-mentioned technical purpose, the online migratory system of a kind of virtual machine provided by the invention comprises:

First module is used for the resource consumption state that the center Control Server is monitored each physical server, is the physical server that target defines the virtual machine (vm) migration demand to optimize the resource consumption state, is designated as migration source physical server;

Second module, be used for choosing candidate's virtual machine to be migrated at migration source physical server, and choose the candidate for each candidate virtual machine to be migrated and move the target physical server, the candidate moves that the target physical server need satisfy its available CPU nuclear number and memory size is enough to the virtual machine use to be migrated for the candidate, with candidate's virtual machine to be migrated and candidate move the target physical server one by one corresponding formation candidate move scheme;

Second module is used for estimating the overall performance interference that each candidate's migration scheme causes, and specifically comprises following submodule:

First submodule, it is more big and transit time is more long to be used for network I/O bandwidth contention according to migration source physical server and target physical server, then transition process disturbs more big criterion to the migration performance that migration source physical server and target physical server cause, and the migration performance that evaluate candidate migration scheme causes disturbs;

Second submodule, be used for after migration, disturbing more big to the same machine performance that migration target physical server causes in conjunction with the more big then virtual machine to be migrated of network I/O resources occupation rate of virtual machine to be migrated, and the demand of the buffer memory of virtual machine to be migrated and memory bandwidth is more big, and the supply of buffer memory and memory bandwidth is more little on the migration target physical server, then virtual machine to be migrated disturbs more big criterion to the same machine performance that migration target physical server causes after migration, and the same machine performance that evaluate candidate migration scheme causes is disturbed;

The 3rd submodule, the migration performance that causes for comprehensive evaluation candidate migration scheme disturbs and disturbs with the machine performance, obtains the overall performance interference that candidate's migration scheme causes;

Four module is used for choosing overall performance from each candidate's migration scheme and disturbs reckling as final migration scheme;

The 5th module is used for migration source physical server and finishes the online migration of virtual machine according to final migration scheme.

Further, in described second module, the virtual machine network I/O bandwidth consumption on the described network I/O bandwidth contention employing physical server and the network of physical server interrupt supply and demand ratio comprehensively to be weighed; Described transit time adopts the memory size of virtual machine to be migrated and the dirty page or leaf of the internal memory rate measurement of virtual machine to be migrated comprehensively to weigh; Described network I/O resources occupation rate adopts network I/O bandwidth or network I/O interruption times or network TCP socket number to weigh; The buffer memory of described virtual machine to be migrated and the demand of memory bandwidth adopt virtual machine cache miss rate and memory usage sum to weigh; The supply of buffer memory and memory bandwidth adopts the cache hit rate of this physical server and internal memory surplus ratio sum to weigh on the described migration target physical server.

Further, described first submodule is used for the specific implementation that migration performance that evaluate candidate migration scheme causes disturbs and is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; The candidate moves scheme＜V _l, P _dThe migration performance that causes disturbs

, ω _sBe virtual machine V to be migrated _lThe migration performance that in transition process migration source physical server is caused disturbs ω _dBe virtual machine V to be migrated _lIn transition process to migration target physical server P _dThe migration performance that causes disturbs, ω _SmThe transition process that moves scheme for all candidates disturbs maximal value, ω to the migration performance that migration source physical server causes _DmThe transition process that moves scheme for all candidates disturbs maximal value to the migration performance that migration target physical server causes;

Wherein, the migration performance that in transition process migration source or target physical server caused of virtual machine to be migrated disturbs

, α is the number of the virtual machine of migration source or the carrying of target physical server, μ _IhBe the average of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, σ _IhBe the standard deviation of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, μ _ItBe the average of the network interruption times of i virtual machine of migration source or target physical server, β _lBe the internal memory after the virtual machine normalization to be migrated, D _lBe the dirty page or leaf of the internal memory of virtual machine to be migrated rate, μ _tAverage for the network interruption times of migration source or target physical server;

The specific implementation that the same machine performance that the evaluate candidate migration scheme that is used for described second submodule causes is disturbed is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; Virtual machine V to be migrated _lSame machine performance disturb

, γ _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe network bandwidth performance that causes is disturbed, ε _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe buffer memory that causes and the performance of memory bandwidth are disturbed, γ _mFor the virtual machine to be migrated in all candidate's migration schemes disturbs maximal value, ε at the network I/O that migration target physical server causes _mThe virtual machine to be migrated that moves scheme for all candidates disturbs maximal value in buffer memory and memory bandwidth performance that migration target physical server causes;

Wherein, ${\mathrm{\γ}}_{\mathrm{ld}}=\frac{{\mathrm{\μ}}_{\mathrm{lt}}}{{\mathrm{\μ}}_{\mathrm{dt}}}(\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_{\mathrm{dt}}}),$ ${\mathrm{\ϵ}}_{\mathrm{ld}}=\frac{{\mathrm{VU}}_{\mathrm{ml}}+{\mathrm{VCM}}_l}{(1-{\mathrm{PU}}_{\mathrm{md}})+(1-{\mathrm{PCM}}_d)},$

μ _LtBe virtual machine V to be migrated _lNetwork interruption times average, μ _DtBe the network interruption average of migration target physical server, α _dBe the virtual machine number of migration target physical server carrying, VU _MlBe virtual machine V to be migrated _lMemory usage, μ _ItBe the average of the network interruption times of migration target physical server i virtual machine, VU _MlBe virtual machine V to be migrated _lMemory usage, VCM _lBe virtual machine V to be migrated _lThe cache miss rate, PU _MdBe the memory usage of migration target physical server, PCM _dCache miss rate for migration target physical server;

The migration performance that the comprehensive candidate's migration scheme that is used for described the 3rd submodule causes disturbs and disturbs with the machine performance and obtains the specific implementation that overall performance that candidate's migration scheme causes disturbs and be: overall performance disturbs T=ρ M+ (1-ρ) N, weight coefficient 0≤ρ≤1, the migration performance that M causes for candidate's migration scheme disturbs, and the same machine performance that N causes for candidate's migration scheme is disturbed.

By the above technical scheme that the present invention conceives, compared with prior art, the present invention has following beneficial effect:

1, the virtual machine performance disturbs evaluation and test comprehensively accurately: the present invention takes all factors into consideration two kinds of performances and disturbs, comprise in the virtual machine (vm) migration process that the performance that causes at source physical server and target physical server disturbs, with and migration finish the same machine performance interference of back on the target physical server.Simultaneously, when estimating the performance interference, this programme is taken all factors into consideration the various dimensions resource consumption situation of virtual machine, comprises the operating position of buffer memory, internal memory and the network bandwidth.

2, be simple and easy to realize: the present invention need not by moving the method for virtual machine load or off-line measurement in advance, obtains the performance of being brought by virtual machine (vm) migration and disturbs, and disturbs but adopt a kind of online mode to calculate online migration performance in real time.The present invention simultaneously need not change the administrative mechanism of existing virtual data center, and with the online moving method compatibility of existing virtual machine.

Description of drawings

Fig. 1 is the online migration process flow diagram of virtual machine that virtualized environment neutrality of the present invention can be disturbed perception.

Fig. 2 is the particular flow sheet of step in the inventive method (3).

Embodiment

In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing and exemplifying embodiment, the present invention is further elaborated.Should be appreciated that concrete exemplifying embodiment described herein only in order to explaining the present invention, and be not used in restriction the present invention.

At first, introduce and use data center of the present invention scene, it relates to a center Control Server and Duo Tai physical server, carries a plurality of virtual machines on every physical server.The center Control Server is responsible for and the communicating by letter of physical server, virtual machine, and finishes the migration decision-making of virtual machine, and physical server is finished virtual machine (vm) migration according to moving the result of decision.

As shown in Figure 1, the various dimensions resource performance disturbs the online moving method of virtual machine of perception may further comprise the steps in the virtualized environment of the present invention:

Whether (1) the resource consumption state of center Control Server (Central Server is called for short CS) monitors physical server (Physical Machine is called for short PM) needs migration to judge virtual machine (Virtual Machine is called for short VM); If need migration, then enter step (2), else process finishes.

The purpose on this step ground is to detect the too high or too low physical server of resource utilization, in order to the virtual machine (vm) migration that moves on the follow-up physical server that resource utilization is too high to other physical server, to reach the purpose of load balancing; Or the virtual machine that moves on the physical server that resource utilization is low excessively is integrated into other servers, to reach the purpose of energy efficient.

Particularly, in current monitoring period, each virtual machine of center Control Server and physical server and carrying thereof communicates, know the resource consumption state of each physical server with the various dimensions resource consumption information of obtaining virtual machine and physical server, and judge whether to need the migration virtual machine come management maintenance data center with this, to optimize the resource consumption state.Wherein resource consumption information specifically comprises: virtual machine V _iThe average μ of network I/O bandwidth consumption _IhAnd standard deviation sigma _Ih, the average μ of network interruption times _It, the dirty page or leaf of internal memory rate D _i, memory usage VU _Mi, and buffer memory (cache) miss rate VCM _iPhysical server P _jThe average μ of network interruption times _Jt, CPU usage PU _Cj, memory usage PU _Mj, and cache miss rate PCM _j

The resource consumption state specifically can adopt CPU usage, memory usage, network I/O bandwidth consumption to wait to weigh.Because CPU usage is to weigh the key index of physical server performance, so the preferred CPU usage of the present invention weighs the resource consumption state, and with this as the adequate condition that triggers virtual machine (vm) migration, elaborate below.Judge whether that the real needs of moving virtual machine are made as: as certain physical server P _jOverload, i.e. CPU usage PU _CjThe resource utilization higher limit, then need to move the virtual machine that moves on this server, to reach the purpose of load balancing; Perhaps as certain physical server P _jResource utilization low excessively, i.e. CPU usage PU _Cj＜resource utilization lower limit, the virtual machine that then needs to move on this server is integrated into other servers, to reach the purpose of energy efficient.Resource utilization higher limit and lower limit are empirical value, and generally speaking, resource utilization higher limit span is 80%～99%, and resource utilization lower limit span is 1%～20%.In general, if many physical servers overloads occur, choose then that to transship severe patient be migration source physical server; If it is low excessively many physical server utilizations of resources to occur, then choose the lowest as migration source physical server; If occur overload physical server and the low excessively physical server of resource utilization simultaneously, consider the overload physical server because the excessive utilization of resource becomes the performance bottleneck of data center easily, and the low excessively physical server of resource utilization just causes the wasting of resources but is very little to the performance impact at data center, and therefore preferred overload physical server is as migration source physical server.

(2) on the physical server of migration source, the center Control Server is chosen a plurality of candidates virtual machine to be migrated and corresponding a plurality of migration target physical servers;

The selection of candidate's virtual machine to be migrated: for the migration source physical server of overload, virtual machine that can it is all is as candidate's virtual machine to be migrated, but it causes calculated amount big, thus the preferred therefrom selected part of the present invention to the bigger virtual machine of physical server resources occupation rate as candidate's virtual machine to be migrated; For the low excessively physical server of resource utilization, then that it is all virtual machines are as candidate's virtual machine to be migrated.

The candidate moves the selection of target physical server: all are transshipped physical servers physical server in addition as alternative physical server.At each virtual machine to be migrated, choose the physical server that is suitable as its migration purpose from alternative physical server, namely the center Control Server is chosen according to the constraint condition of migration target physical server (being that resources such as the available core cpu number of target physical server and memory size must can be enough to for moving the virtual machine use).Then according to virtual machine to be migrated and target physical server one by one corresponding structure obtain the candidate and move scheme, namely to candidate virtual machine V _lFind out Z candidate and move target physical server { P _l..., P _Z, make up the candidate and move scheme＜V _l, P _d, d=1,2 ..., Z.

(3) for each the candidate is moved scheme, the center Control Server calculates the overall performance interference value that this migration scheme causes the data center.

Particularly, for each the candidate is moved tuple＜V _l, P _d(representative is with virtual machine V _lMigrate to server P _d), the center Control Server is with this candidate virtual machine V to be migrated _l, the candidate moves source physical server P _s(namely carry V _lPhysical server) and candidate migration and target physical server P _dVarious dimensions resource consumption value, calculate corresponding performance and disturb discreet value.

(4) in all candidate lists (scheme), the center Control Server is selected the migration scheme (virtual machine to be migrated and migration target physical server) of overall performance interference value minimum.

(5) the source physical server utilizes the online migration interface that Virtual Machine Manager program (Hypervisor) provides, with the online target physical server that migrates to of selected virtual machine to be migrated.

Step (3) is taken all factors into consideration the factor of various dimensions resource consumptions such as buffer memory, memory bandwidth, network I/O bandwidth, estimates among the virtual machine (vm) migration process and the performance of bringing after migration interference comprehensively.As shown in Figure 2, at each the candidate is moved scheme, step (3) is estimated the performance interference that this candidate's scheme of moving produces according to following substep:

(3-1) transition process of center Control Server evaluation migration scheme correspondence disturbs the migration performance that migration source physical server and target physical server in this scheme cause.

The technical thought of this step: because virtual machine (vm) migration relates generally to the consumption to the network bandwidth resources of migration source physical server and target physical server, therefore migration is disturbed and is presented as that mainly migration source physical server and target physical server network I/O performance disturb.Furthermore, the virtual machine (vm) migration process is to network I/O bandwidth contention (when the virtual machine network I/O bandwidth aggregate demand of carrying when a physical server surpasses the network I/O bandwidth of this physical server, namely producing virtual machine network I/O bandwidth contention) performance that produces of comparatively serious physical server disturbs greatlyyer, and disturbs less to the performance that the lighter physical server of network I/O competition produces; The virtual machine (vm) migration required time is more long, it disturbs more big to the performance that migration source physical server and target physical server produce, the virtual machine (vm) migration required time is more short, and it disturbs more little to the performance that migration source physical server and target physical server produce.To sum up, the migration performance on the source and destination physical server disturbs virtual machine network I/O competition intensity and the function of virtual machine (vm) migration time (for example product function) that can be expressed as on the physical server.

In the concrete evaluation procedure, can choose suitable parameter and weigh virtual machine network I/O bandwidth contention intensity and virtual machine (vm) migration time and just can quantize migration performance and disturb.

The measurement parameter of virtual machine network I/O bandwidth contention intensity is chosen and can be chosen the network I/O interruption times of virtual machine and physical server and weigh according to the I/O scheduling mechanism of Intel Virtualization Technology theory.Among the present invention, virtual machine network I/O bandwidth contention intensity on the physical server is weighed by two factors, one is the fluctuating range of the virtual machine network I/O bandwidth on the physical server, allocation scheme according to virtual machine I/O bandwidth resources, under I/O bandwidth resources situation in short supply, the big I of virtual machine network I/O fluctuating range directly reflects virtual machine network I/O bandwidth contention intensity; Another is the network interruption supply and demand ratio of physical server, and the supply and demand ratio of resource can be used to reflect the intensity of resource contention, and supply and demand ratio is more big, shows that then competition is more fierce.This is because these two factors have reason, therefore can weigh virtual machine network I/O bandwidth contention intensity.More than the value of two factors more big, show that the network I/O bandwidth contention intensity of virtual machine on the physical server is more big.In the present invention, according to statistical theory, the fluctuating range of separate unit virtual machine network I/O bandwidth can be weighed by the coefficient of variation (being the ratio of I/O bandwidth criteria difference and I/O bandwidth average) of this virtual machine network I/O bandwidth; In addition, according to the marketing economics theory, the network of physical server interrupts supply and demand ratio can be weighed divided by the network interruption times (supply) that this physical server provides by the virtual machine network interruption times summation (demand) on this physical server.Above measurement mode is simple, calculated amount is little, and all parameters can onlinely be obtained, and can accurately weigh above two indexs.

The measurement parameter of virtual machine (vm) migration time is chosen can be according to Intel Virtualization Technology theory and existing correlative study, choose the dirty page or leaf of the internal memory rate of the memory size of virtual machine to be migrated, virtual machine to be migrated, the measurements such as network I/O amount of bandwidth of physical server, specifically, the virtual machine (vm) migration time is directly proportional with preceding two parameters, be inversely proportional to a back parameter, concrete virtual machine (vm) migration Time Calculation model is comparatively complicated.The virtual machine (vm) migration time of the present invention has been done rational simplification, also choosing wherein the product of two aspect factors weighs, one is the memory size of virtual machine to be migrated, and another is the dirty page or leaf of the internal memory of virtual machine to be migrated rate, and this has low, the technique effect accurately of computation complexity.

Particularly, calculate also regularization candidate virtual machine V to be migrated _lAt migration source physical server P _sWith migration target physical server P _dOn the migration interference value M that produces.In the preferred mode of the present invention, M can be tried to achieve by following formula:

$M=\frac{{\mathrm{\ω}}_s}{{\mathrm{\ω}}_{\mathrm{sm}}}+\frac{{\mathrm{\ω}}_d}{{\mathrm{\ω}}_{\mathrm{dm}}}$

Wherein, ω _sWith ω _dBe respectively virtual machine V _iIn transition process to source physical server P _sWith target physical server P _dThe migration interference value that causes.ω _SmWith ω _DmBe respectively the maximal value of disturbing in migration source physical server and the migration on the migration target physical server in whole candidate's migration schemes.ω _sWith ω _dCan be tried to achieve by the collected resource consumption information of step (1) center Control Server, concrete account form is as follows:

The migration performance that virtual machine to be migrated causes migration source or target physical server in transition process disturbs

$\mathrm{\ω}=(\underset{i=1}{\overset{\mathrm{\α}}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{\mathrm{\α}}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_t}){\mathrm{\β}}_l{D}_l$

Wherein, α is the number of the virtual machine of migration source or the carrying of target physical server, μ _IhBe the average of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, σ _IhBe the standard deviation of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, μ _ItBe the average of the network interruption times of i virtual machine of migration source or target physical server, β _lBe the internal memory after the virtual machine normalization to be migrated, D _lBe the dirty page or leaf of the internal memory of virtual machine to be migrated rate, μ _tAverage for the network interruption times of migration source or target physical server.

(3-2) the same machine of center Control Server calculated candidate virtual machine to be migrated on migration target physical server disturbs.

The technical thought of this step: virtual machine disturbs with machine and is mainly weighed by two aspect factors, the network I/O performance that to be virtual machine to be migrated cause on migration target physical server is disturbed, and another is that the buffer memory that causes on migration target physical server of virtual machine to be migrated and the performance of memory bandwidth are disturbed.On the one hand, the network I/O resources occupation rate more big (proportional relation) of virtual machine to be migrated, it disturbs more serious to the network I/O performance that migration target physical server causes.Therefore, the network I/O performance that causes on migration target physical server of virtual machine to be migrated is disturbed network I/O resources occupation rate (specifically weigh parameter and can be network I/O bandwidth, network I/O interruption times, the network TCP socket(socket) number that can be expressed as virtual machine to be migrated etc.) function.Concrete measurement mode can be distributed correlation theory according to resource in the Intel Virtualization Technology, chooses the network I/O resource requirement size of virtual machine to be migrated, the factors such as virtual machine network I/O bandwidth contention intensity on the migration target physical server.Consider that computation complexity is little and do not lose the advantage of accuracy, the present invention adopts linear equation to weigh this performance and disturbs, the slope of this linear equation is the ratio of the virtual machine network I/O bandwidth contention intensity on the migration target physical server with the network I/O interruption times of migration target physical server, the network I/O resources occupation rate of virtual machine to be migrated is chosen to be virtual machine network I/O interruption times, because network I/O interruption times can reflect the network I/O resource requirement size of virtual machine.Wherein, the virtual machine network I/O bandwidth contention intensity on the migration target physical server can be tried to achieve by step (3-1).

On the other hand, the buffer memory of virtual machine to be migrated and the demand of memory bandwidth more big (proportional relation), the supply more little (inverse relation) of buffer memory and memory bandwidth on the migration target physical server shows that then buffer memory that virtual machine to be migrated causes and the performance of memory bandwidth disturb more serious on migration target physical server.Its concrete measurement mode can be chosen virtual machine to be migrated and the cache miss rate of migration target physical server, virtual machine to be migrated and the parameters such as memory usage that move the target physical server according to the correlation theory of resource distribution in the Intel Virtualization Technology.Consider that computation complexity is little and do not lose the advantage of accuracy, the present invention adopts the resource supply and demand ratio in the marketing economics theory, weighs buffer memory that virtual machine to be migrated causes and the performance of memory bandwidth and disturb on migration target physical server.Wherein, the demand of the buffer memory of virtual machine and memory bandwidth can be weighed by this virtual machine cache miss rate and memory usage sum in the current period; The buffer memory of physical server and the supply of memory bandwidth can be weighed by this physical server cache hit rate (1-miss rate) in the current period and internal memory surplus ratio (1-occupancy) sum.

Particularly, calculate also regularization candidate virtual machine V to be migrated _lAt migration target physical server P _dOn the same machine interference value N that produces.In the preferred mode of the present invention, N can be tried to achieve by following formula:

$N=\frac{{\mathrm{\γ}}_{\mathrm{ld}}}{{\mathrm{\γ}}_m}+\frac{{\mathrm{\ϵ}}_{\mathrm{ld}}}{{\mathrm{\ϵ}}_m}$

Wherein, γ _LdBe virtual machine V to be migrated _lAt migration target physical server P _dOn the network I/O performance interference value that causes, ε _LdBe virtual machine V to be migrated _lAt migration target physical server P _dOn the buffer memory that causes and the performance interference value of memory bandwidth.γ _mWith ε _mBe respectively the network I/O interference of virtual machine to be migrated on migration target physical server and the maximal value of buffer memory and the interference of memory bandwidth performance in whole candidate's migration schemes.γ _LdWith ε _LdCan be tried to achieve by the collected resource consumption information of step (1) center Control Server, specifically according to following formula:

${\mathrm{\γ}}_{\mathrm{ld}}=\frac{{\mathrm{\μ}}_{\mathrm{lt}}}{{\mathrm{\μ}}_{\mathrm{dt}}}(\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_{\mathrm{dt}}}),$ ${\mathrm{\ϵ}}_{\mathrm{ld}}=\frac{{\mathrm{VU}}_{\mathrm{ml}}+{\mathrm{VCM}}_l}{(1-{\mathrm{PU}}_{\mathrm{md}})+(1-{\mathrm{PCM}}_d)}.$

Calculate γ _LdFormula in, μ _LtBe virtual machine V to be migrated _lNetwork interruption times average, μ _DtBe migration target physical server P _dNetwork interrupt average, α _dBe migration target physical server P _dThe virtual machine number of carrying.

Calculate ε _LdFormula in, VU _MlBe virtual machine V to be migrated _lMemory usage, VCM _lBe virtual machine V to be migrated _lThe cache miss rate, PU _MdBe migration target physical server P _dMemory usage (0＜PU _Md＜1), PCM _dBe migration target physical server P _dCache miss rate (0＜PCM _d＜1)

(3-3) the center Control Server merges the performance interference value in above two substeps, to estimate the overall performance interference value that virtual machine is caused with migration afterwards in transition process.Particularly, in the virtual machine (vm) migration process and among the migration overall performance interference value T afterwards, the shared weight size of same machine performance interference N after the center Control Server can utilize the dynamic adjustment of weight coefficient ρ to be disturbed M and moved by the performance in the virtual machine (vm) migration process.Can be tried to achieve by following formula with migration overall performance interference value T afterwards in the virtual machine (vm) migration process:

T＝ρM+(1-ρ)N

Wherein, the weight size that parameter ρ implication disturbs shared overall performance to disturb for migration, its span is 0≤ρ≤1.Performance is disturbed adjustable parameter ρ, can be for data center administrator by its actual demand, dynamically adjust the weight size of moving interference and disturbing with machine in the online moving method of virtual machine, make this invention when the performance interference of the online migration of evaluation and test virtual machine, have higher flexibility.Particularly, the running load of dissimilar (network I/O intensity, CPU computation-intensive, internal memory computation-intensives) disturbs migration and is different with the sensitivity of machine interference.In general, the intensive load of network I/O is disturbed more responsive to migration, if the running load of migration source and target physical server is network I/O intensity, the ρ value needs greater than 0.5; And the load of CPU computation-intensive is more responsive to disturbing with machine, if the running load of migration source and target physical server is the CPU computation-intensive, the ρ value needs less than 0.5.If the running load of the device of migration source and target physical service is the mixed type load, general ρ value is 0.5.Therefore, the center Control Server can be adjusted as required based on the running load type of virtual machine on migration source and the target physical server.

Above scheme has been described the required step of virtual machine (vm) migration of data center's single performance interference perception.Yet the single virtual machine (vm) migration may not can reach targets such as the desired load balancing of center Control Server or energy consumption saving; In addition, because the running load of data center fluctuates in time, cause virtual machine (vm) migration physical server afterwards to occur resource utilization overload or low excessively situation again.Under above two kinds of situations, the present invention can finish repeatedly necessary virtual machine (vm) migration by the physical server resource utilization of center Control Server regular monitoring data center, with maintenance, management data center.

Adopt the average of the interior virtual machine of current period and physical server resource utilization sequence as the input computing parameter of estimating each scheme to be migrated in the step (3) among the present invention, the method is simple, expense is little, feasibility is high.Can adopt other comparatively complicated, accurate computing method to the expansion of this programme, such as statistical methods such as the maximal value of getting resource utilization sequence in the current period or employing autoregressive model, Hidden Markov Model (HMM).

Application example

For the feasibility and the validity that embody this programme, the prototype system of building under true environment is verified the present invention.This system comprises 1 center Control Server, 9 physical servers, and 1 network file system server.Particularly, every physical server configuration two-way four nuclear Intel Xeon E56202.40GHz processors, 12MB L3 level shared buffer memory, the 24GB internal memory, 1Gbps full duplex network adapter, physical server operation CentOS5.5 operating system (kernel of Linux2.6.18.8-Xen4.1.1), and all servers are connected to the network switch of 1Gbps, in the LAN (Local Area Network) that coexists.Virtual machine manager adopts the half virtual mode of Xen4.1.1, and the disk mirroring file of all virtual machines is stored in the network file system server, by the mode carry virtual machine image of network.Prototype system is carried 50 virtual machines altogether, the load that per 10 virtual machines operation is a type, and 5 kinds of virtual machine load concrete configurations are as shown in table 1:

The center Control Server is regularly collected the resource consumption information of each physical server and virtual machine, and in this test macro, the information collection cycle of center Control Server is set at 10 seconds, and migration is disturbed with the scale parameter ρ that disturbs with machine and is made as 0.5.

Test experiments is the result show: under the situation of a virtual machine of migration, respectively with the virtual machine (vm) migration scheme First-Fit Decreasing(FFD that with the energy consumption saving is target), and be that the virtual machine (vm) migration scheme Sandpiper of target compares with the load balancing, this programme can improve netperf(network-intensive type) performance of load is up to 55%-70%, improve Hadoop, NASA parallel benchmark and SPECWeb(CPU, internal memory, the medium consumption-type of network) performance of load reaches 10%-18%.Simultaneously, this programme is only introduced small CPU expense and network I/O expense.Specifically, the network I/O expense that this programme produced in the Control Server of center in per 10 seconds only is 14400 byte numbers, to operation SPECCPU(CPU intensity in the virtual machine) influence of load only is about 1 second.

Those skilled in the art will readily understand; the above only is preferred embodiment of the present invention; not in order to limiting the present invention, all any modifications of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., all should be included within protection scope of the present invention.

Claims (10)

1. online moving method of virtual machine may further comprise the steps:

(1) the center Control Server is monitored the resource consumption state of each physical server, is the physical server that target defines the virtual machine (vm) migration demand to optimize the resource consumption state, is designated as migration source physical server;

(2) choose candidate's virtual machine to be migrated at migration source physical server, and choose the candidate for each candidate virtual machine to be migrated and move the target physical server, the candidate moves that the target physical server need satisfy its available CPU nuclear number and memory size is enough to the virtual machine use to be migrated for the candidate, with candidate's virtual machine to be migrated and candidate move the target physical server one by one corresponding formation candidate move scheme;

(3) estimate the overall performance interference that each candidate's migration scheme causes, its process is:

More big and transit time is more long according to the network I/O bandwidth contention of migration source physical server and target physical server, then transition process disturbs more big criterion to the migration performance that migration source physical server and target physical server cause, and the migration performance that evaluate candidate migration scheme causes disturbs;

The more big then virtual machine to be migrated of network I/O resources occupation rate in conjunction with virtual machine to be migrated disturbs more big to the same machine performance that migration target physical server causes after migration, and the demand of the buffer memory of virtual machine to be migrated and memory bandwidth is more big, and the supply of buffer memory and memory bandwidth is more little on the migration target physical server, then virtual machine to be migrated disturbs more big criterion to the same machine performance that migration target physical server causes after migration, and the same machine performance that evaluate candidate migration scheme causes is disturbed;

The migration performance that comprehensive evaluation candidate's migration scheme causes disturbs and disturbs with the machine performance, obtains the overall performance interference that candidate's migration scheme causes;

(4) choosing overall performance from each candidate's migration scheme disturbs reckling as final migration scheme;

(5) migration source physical server is finished the online migration of virtual machine according to final migration scheme.

2. the online moving method of virtual machine according to claim 1, it is characterized in that, in the described step (3), the virtual machine network I/O bandwidth consumption on the described network I/O bandwidth contention employing physical server and the network of physical server interrupt supply and demand ratio comprehensively to be weighed; Described transit time adopts the memory size of virtual machine to be migrated and the dirty page or leaf of the internal memory rate measurement of virtual machine to be migrated comprehensively to weigh; Described network I/O resources occupation rate adopts network I/O bandwidth or network I/O interruption times or network TCP socket number to weigh; The buffer memory of described virtual machine to be migrated and the demand of memory bandwidth adopt virtual machine cache miss rate and memory usage sum to weigh; The supply of buffer memory and memory bandwidth adopts the cache hit rate of this physical server and internal memory surplus ratio sum to weigh on the described migration target physical server.

3. the online moving method of virtual machine according to claim 1 and 2 is characterized in that,

The specific implementation that the migration performance that evaluate candidate migration scheme causes in the described step (3) disturbs is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; The candidate moves scheme＜V _l, P _dThe migration performance that causes disturbs , ω _sBe virtual machine V to be migrated _lThe migration performance that in transition process migration source physical server is caused disturbs ω _dBe virtual machine V to be migrated _lIn transition process to migration target physical server P _dThe migration performance that causes disturbs, ω _SmThe transition process that moves scheme for all candidates disturbs maximal value, ω to the migration performance that migration source physical server causes _DmThe transition process that moves scheme for all candidates disturbs maximal value to the migration performance that migration target physical server causes;

Wherein, the migration performance that in transition process migration source or target physical server caused of virtual machine to be migrated disturbs , α is the number of the virtual machine of migration source or the carrying of target physical server, μ _IhBe the average of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, σ _IhBe the standard deviation of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, μ _ItBe the average of the network interruption times of i virtual machine of migration source or target physical server, β _lBe the internal memory after the virtual machine normalization to be migrated, D _lBe the dirty page or leaf of the internal memory of virtual machine to be migrated rate, μ _tAverage for the network interruption times of migration source or target physical server;

The specific implementation that the same machine performance that described step (3) evaluate candidate migration scheme causes is disturbed is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; Virtual machine V to be migrated _lSame machine performance disturb

, γ _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe network bandwidth performance that causes is disturbed, ε _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe buffer memory that causes and the performance of memory bandwidth are disturbed, γ _mFor the virtual machine to be migrated in all candidate's migration schemes disturbs maximal value, ε in the network bandwidth performance that migration target physical server causes _mThe virtual machine to be migrated that moves scheme for all candidates disturbs maximal value in buffer memory and memory bandwidth performance that migration target physical server causes;

Wherein, ${\mathrm{\γ}}_{\mathrm{ld}}=\frac{{\mathrm{\μ}}_{\mathrm{lt}}}{{\mathrm{\μ}}_{\mathrm{dt}}}(\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_{\mathrm{dt}}}),$ ${\mathrm{\ϵ}}_{\mathrm{ld}}=\frac{{\mathrm{VU}}_{\mathrm{ml}}+{\mathrm{VCM}}_l}{(1-{\mathrm{PU}}_{\mathrm{md}})+(1-{\mathrm{PCM}}_d)},$

μ _LtBe virtual machine V to be migrated _lNetwork interruption times average, μ _DtBe the network interruption average of migration target physical server, α _dBe the virtual machine number of migration target physical server carrying, VU _MlBe virtual machine V to be migrated _lMemory usage, μ _ItBe the average of the network interruption times of migration target physical server i virtual machine, VU _MlBe virtual machine V to be migrated _lMemory usage, VCM _lBe virtual machine V to be migrated _lThe cache miss rate, PU _MdBe the memory usage of migration target physical server, PCM _dCache miss rate for migration target physical server;

The migration performance that the comprehensive candidate's migration scheme of described step (3) causes disturbs and disturbs with the machine performance and obtains the specific implementation that overall performance that candidate's migration scheme causes disturbs and be: overall performance disturbs T=ρ M+ (1-ρ) N, weight coefficient 0≤ρ≤1, the migration performance that M causes for candidate's migration scheme disturbs, and the same machine performance that N causes for candidate's migration scheme is disturbed.

4. according to claim 1 or the online moving method of 2 or 3 described virtual machines, it is characterized in that if the running load of migration source physical server and target physical server is network I/O intensity, the ρ value is greater than 0.5; If the running load of migration source physical server and target physical server is the CPU computation-intensive, the ρ value is less than 0.5; If the running load of the device of migration source and destination physics service is the mixed type load of network I/O intensity and CPU computation-intensive, ρ value 0.5.

5. according to claim 1 or the online moving method of 2 or 3 described virtual machines, it is characterized in that, described step (1) is specially: the center Control Server is monitored the resource consumption state of each physical server, determines the physical server that resource utilization is too high or too low according to the resource consumption state; If exist many physical server resource utilizations too high, choose then that to transship severe patient be migration source physical server; If exist many physical server utilizations of resources low excessively, then choose the lowest as migration source physical server; If the too high physical server of resource utilization and the low excessively physical server of resource utilization occur simultaneously, then choose the too high physical server of resource utilization as migration source physical server; Described resource consumption state adopts CPU usage or memory usage or network I/O bandwidth consumption to weigh.

6. the online moving method of virtual machine according to claim 5, it is characterized in that, the specific implementation of choosing candidate's virtual machine to be migrated at migration source physical server is: for the too high migration source physical server of resource utilization, selected part to the big virtual machine of physical server resources occupation rate as candidate's virtual machine to be migrated; For the low excessively migration source physical server of resource utilization, the virtual machine that it is all is as candidate's virtual machine to be migrated.

7. the online migratory system of virtual machine is characterized in that, comprising:

First module is used for the resource consumption state that the center Control Server is monitored each physical server, is the physical server that target defines the virtual machine (vm) migration demand to optimize the resource consumption state, is designated as migration source physical server;

Second module, be used for choosing candidate's virtual machine to be migrated at migration source physical server, and choose the candidate for each candidate virtual machine to be migrated and move the target physical server, the candidate moves that the target physical server need satisfy its available CPU nuclear number and memory size is enough to the virtual machine use to be migrated for the candidate, with candidate's virtual machine to be migrated and candidate move the target physical server one by one corresponding formation candidate move scheme;

Second module is used for estimating the overall performance interference that each candidate's migration scheme causes, and specifically comprises following submodule:

First submodule, it is more big and transit time is more long to be used for network I/O bandwidth contention according to migration source physical server and target physical server, then transition process disturbs more big criterion to the migration performance that migration source physical server and target physical server cause, and the migration performance that evaluate candidate migration scheme causes disturbs;

Second submodule, be used for after migration, disturbing more big to the same machine performance that migration target physical server causes in conjunction with the more big then virtual machine to be migrated of network I/O resources occupation rate of virtual machine to be migrated, and the demand of the buffer memory of virtual machine to be migrated and memory bandwidth is more big, and the supply of buffer memory and memory bandwidth is more little on the migration target physical server, then virtual machine to be migrated disturbs more big criterion to the same machine performance that migration target physical server causes after migration, and the same machine performance that evaluate candidate migration scheme causes is disturbed;

The 3rd submodule, the migration performance that causes for comprehensive evaluation candidate migration scheme disturbs and disturbs with the machine performance, obtains the overall performance interference that candidate's migration scheme causes;

Four module is used for choosing overall performance from each candidate's migration scheme and disturbs reckling as final migration scheme;

The 5th module is used for migration source physical server and finishes the online migration of virtual machine according to final migration scheme.

8. the online migratory system of virtual machine according to claim 7, it is characterized in that, in described second module, the virtual machine network I/O bandwidth consumption on the described network I/O bandwidth contention employing physical server and the network of physical server interrupt supply and demand ratio comprehensively to be weighed; Described transit time adopts the memory size of virtual machine to be migrated and the dirty page or leaf of the internal memory rate measurement of virtual machine to be migrated comprehensively to weigh; Described network I/O resources occupation rate adopts network I/O bandwidth or network I/O interruption times or network TCP socket number to weigh; The buffer memory of described virtual machine to be migrated and the demand of memory bandwidth adopt virtual machine cache miss rate and memory usage sum to weigh; The supply of buffer memory and memory bandwidth adopts the cache hit rate of this physical server and internal memory surplus ratio sum to weigh on the described migration target physical server.

9. the online migratory system of virtual machine according to claim 8 is characterized in that,

The specific implementation that the migration performance that evaluate candidate migration scheme causes in described first submodule disturbs is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; The candidate moves scheme＜V _l, P _dThe migration performance that causes disturbs

ω _sBe virtual machine V to be migrated _lThe migration performance that in transition process migration source physical server is caused disturbs ω _dBe virtual machine V to be migrated _lIn transition process to migration target physical server P _dThe migration performance that causes disturbs, ω _SmThe transition process that moves scheme for all candidates disturbs maximal value, ω to the migration performance that migration source physical server causes _DmThe transition process that moves scheme for all candidates disturbs maximal value to the migration performance that migration target physical server causes;

Wherein, the migration performance that in transition process migration source or target physical server caused of virtual machine to be migrated disturbs

, α is the number of the virtual machine of migration source or the carrying of target physical server, μ _IhBe the average of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, σ _IhBe the standard deviation of the network I/O bandwidth consumption of i virtual machine of migration source or target physical server, μ _ItBe the average of the network interruption times of i virtual machine of migration source or target physical server, β _lBe the internal memory after the virtual machine normalization to be migrated, D _lBe the dirty page or leaf of the internal memory of virtual machine to be migrated rate, μ _tAverage for the network interruption times of migration source or target physical server;

The specific implementation that the same machine performance that the described second submodule evaluate candidate migration scheme causes is disturbed is: make candidate's migration scheme be expressed as＜V _l, P _d, V _lBe virtual machine to be migrated, P _dBe migration target physical server; Virtual machine V to be migrated _lSame machine performance disturb

, γ _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe network bandwidth performance that causes is disturbed, ε _LdBe virtual machine V to be migrated _lAt migration target physical server P _dThe buffer memory that causes and the performance of memory bandwidth are disturbed, γ _mFor the virtual machine to be migrated in all candidate's migration schemes disturbs maximal value, ε at the network I/O that migration target physical server causes _mThe virtual machine to be migrated that moves scheme for all candidates disturbs maximal value in buffer memory and memory bandwidth performance that migration target physical server causes;

Wherein, ${\mathrm{\γ}}_{\mathrm{ld}}=\frac{{\mathrm{\μ}}_{\mathrm{lt}}}{{\mathrm{\μ}}_{\mathrm{dt}}}(\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}\frac{{\mathrm{\σ}}_{\mathrm{ih}}}{{\mathrm{\μ}}_{\mathrm{ih}}}+\frac{\underset{i=1}{\overset{{\mathrm{\α}}_d}{\mathrm{\Σ}}}{\mathrm{\μ}}_{\mathrm{it}}}{{\mathrm{\μ}}_{\mathrm{dt}}}),$ ${\mathrm{\ϵ}}_{\mathrm{ld}}=\frac{{\mathrm{VU}}_{\mathrm{ml}}+{\mathrm{VCM}}_l}{(1-{\mathrm{PU}}_{\mathrm{md}})+(1-{\mathrm{PCM}}_d)},$

μ _LtBe virtual machine V to be migrated _lNetwork interruption times average, μ _DtBe the network interruption average of migration target physical server, α _dBe the virtual machine number of migration target physical server carrying, VU _MlBe virtual machine V to be migrated _lMemory usage, μ _ItBe the average of the network interruption times of migration target physical server i virtual machine, VU _MlBe virtual machine V to be migrated _lMemory usage, VCM _lBe virtual machine V to be migrated _lThe cache miss rate, PU _MdBe the memory usage of migration target physical server, PCM _dCache miss rate for migration target physical server;

Migration performance that the described comprehensive candidate's migration of the 3rd submodule scheme causes disturbs and disturbs with the machine performance and obtains the specific implementation that overall performance that candidate's migration scheme causes disturbs and be: overall performance disturbs T=ρ M+ (1-ρ) N, weight coefficient 0≤ρ≤1, the migration performance that M causes for candidate's migration scheme disturbs, and the same machine performance that N causes for candidate's migration scheme is disturbed.

10. according to claim 7 or the online migratory system of 8 or 9 described virtual machines, it is characterized in that if the running load of migration source physical server and target physical server is network I/O intensity, the ρ value is greater than 0.5; If the running load of migration source physical server and target physical server is the CPU computation-intensive, the ρ value is less than 0.5; If the running load of the device of migration source and destination physics service is the mixed type load of network I/O intensity and CPU computation-intensive, ρ value 0.5.

Images