CN103440158B - The hotspot migration method of facing cloud scheduling of resource - Google Patents

Abstract

The invention discloses a kind of hotspot migration method of facing cloud scheduling of resource, causing removing the large technical matters of cost for solving existing hotspot migration method consumes resources.Technical scheme is that the virtual machine in physical machine is carried out principal and subordinate's division, virtual machine in each physical machine is made up of from virtual machine a host virtual machine and several, each service operation is on Master, in order to ensure the normal operation of Master due to business when load too high is collapsed suddenly, each Master to have in several Slave distributions in other physical machine, and regularly and Master carry out data syn-chronization.Each physical machine has Node Controller to collect the resource information on each virtual machine.When Node Controller finds that certain physical machine exists focus, then notify cluster manager dual system, cluster manager dual system judges according to the focus situation in each physical machine, selects suitable physical machine to be moved by focus, thus eliminates focus, decreases migration cost.

Description

The hotspot migration method of facing cloud scheduling of resource

Technical field

The present invention relates to a kind of hotspot migration method, particularly relate to a kind of hotspot migration method of facing cloud scheduling of resource.

Background technology

Intel Virtualization Technology is one of technology foundation stone building cloud computing environment, is that cloud computing technology provides strong realization rate at infrastructure layer.At large-scale data center, the load of virtual machine number and virtual machine often can change with the demand of user and application, virtual machine needs to carry out dynamic resource adjustment, namely will hot point resource timely in removal system, thus reaches the load balancing of whole system.Virtual machine live migration technology is mainly used to the problem of load balancing solving cloud computing system at present.

Document [WOODT.A.Black-Boxandgray-boxstrategiesforvirtualmachinem igration [C] .ProceedingsoftheACMSymposiumonNetworkedSystemsDesignand Implementation, P229-242, April1,2007] disclose a kind of load equilibration scheduling method based on thermophoresis technology.The method devises a load balance scheduler, be made up of Centralized Controller and watch-dog, the CPU of virtual machine, internal memory and network I/O usage statistics are regularly sent to Centralized Controller by watch-dog, Centralized Controller judges where there occurs shortage of resources according to the statistics of each physical machine and virtual machine, then use heuritic approach computation migration regulation scheme, be distributed to watch-dog and implement.The method adopt CPU, internal memory and network three the Volume reciprocal of product of class resource idleness represent the busy extent of physical machine or virtual machine, for it in searching system the physical machine of minimum Volume as migration target, until all focuses are eliminated.The benefit of the method is by unified for multiple resources process, facilitate dispatching algorithm decision-making, but it is nervous to express which resource, may cause from overload physical machine migration walk the virtual machine that an internal memory is large and CPU is not busy.

Summary of the invention

Cause remove the large deficiency of cost to overcome existing hotspot migration method consumes resources, the invention provides a kind of hotspot migration method of facing cloud scheduling of resource.Virtual machine in physical machine is carried out principal and subordinate's division by the method, virtual machine in each physical machine is made up of from virtual machine (Slave) a host virtual machine (Master) and several, each service operation is on Master, in order to ensure the normal operation of Master due to business when load too high is collapsed suddenly, each Master to have in several Slave distributions in other physical machine, and regularly and Master carry out data syn-chronization.Each physical machine has Node Controller to collect the resource information on each virtual machine, as CPU and internal memory behaviour in service etc.When Node Controller finds that certain physical machine exists focus, then notify cluster manager dual system, cluster manager dual system carries out judging (whether exceeding resource threshold) according to the focus situation in each physical machine, selects suitable physical machine to be moved by focus, thus eliminates focus.The method can eliminate the focus of cluster internal fast, reduces migration cost, realizes load balancing rapidly in server cluster inside.

The technical solution adopted for the present invention to solve the technical problems is: a kind of hotspot migration method of facing cloud scheduling of resource, is characterized in comprising the following steps:

Step one, virtual machine define.

For the server of one group of specified configuration, each server has some virtual machines being called tenant, several tenants form one group, distribute a keeper.

Step 2, default.

1) system comprises N platform physical server, SV ₁... SV _n;

2) each tenant there is memory usage with the utilization rate of CPU the CPU use amount of each server and internal memory use amount are the summations of the CPU of all tenants and the use amount of internal memory.That is: ${\mathrm{SV}}_i.\mathrm{mem}=\underset{k=1}{\overset{\mathrm{Ki}}{\mathrm{\Σ}}}{t}_i^K.\mathrm{mem},$ ${\mathrm{SV}}_i.\mathrm{cpu}=\underset{k=1}{\overset{\mathrm{Ki}}{\mathrm{\Σ}}}{t}_i^K.\mathrm{cpu}.$

3) on server, the unit of CPU and internal memory is immutable, and each server has two indices, a memory threshold SV _i.memTh with one CPU threshold value SV _i.cpuTh, this threshold value can not be exceeded, if any one index exceedes on server, claims this server to be a focus.

4) each server has two attributes, memory size SV _i.memCap with CPU capacity SV _i.cpuCap, if SV _i.memCap < 0 or SV _i.cpuCap < 0, this server just becomes hot spot server.Clearly, ${\mathrm{SV}}_i.\mathrm{memCap}={\mathrm{SV}}_i.\mathrm{memTh}-{\mathrm{\&Sigma;}}_{k=1}^{k_i}{t}_i^k.\mathrm{mem},$ And ${\mathrm{SV}}_i.\mathrm{cpuCap}={\mathrm{SV}}_i.\mathrm{cpuTh}-{\mathrm{\&Sigma;}}_{k=1}^{k_i}{t}_i^k.\mathrm{cpu}.$

Step 3, server and virtual machine attribute define.

For each server S V _i, have as properties:

● SV _i.memCap: server S V _iinternal memory residual capacity.

● SV _i.cpuCap: server S V _icPU residual capacity.

● SV _i.memTh: server S V _imemory threshold, exceed then heating point.

● SV _i.memTh: server S V _icPU threshold value, exceed then heating point.

● SV _i.tenants: server S V _ithe set of upper tenant

For each tenant have and each tenant has as properties:

● the identifier of tenant.

● whether tenant is gerentocratic Boolean.

● the memory usage of tenant.

● the CPU usage of tenant.

● the size of the data volume of tenant.

Step 4, basic assumption.

Suppose that X represents any one resource.For N number of numeral { x ₁... x _n, definition mean value variance ${\mathrm{\&sigma;}}_x^2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(\mathrm{xi}-\mathrm{\&mu;x})}^2.$

Therefore, following theorem is obtained:

Theorem 1: if remain unchanged, work as x ₁=x ₂=...=x _ntime minimum.

Theorem 2: for X={x ₁... x _n, obtain:

In formula, be the yardstick measuring resource capacity balance between server, work as x ₁=x ₂=...=x _n=μ _xtime, reach minimum value 0.

In order to the balance between control CPU and internal memory, introduce variable C _memand C _cpu, and be defined as follows:

internal memory correction parameter on individual server.

cPU correction parameter on individual server.

the composed correction parameter of the CPU on individual server and internal memory.

the migration cost of certain tenant.

Step 5, focus remove algorithm.

Algorithm inputs: one group of server S V in cluster _1.... SV _n, wherein may include hot spot server.

Algorithm exports: the strategy set S={ (t, target) of migration }, wherein each strategy is two tuples, comprises the destination server target of a tenant t that will move and its preparation migration.

Initial value: strategy set S=Null; Counter i=0;

Algorithm flow:

Step1: from server list SV _1.... SV _nin obtain i server, judge its memory size SV _ior CPU capacity SV .memCap _i.cpuCap whether be less than 0, if so, then this server is hot spot server;

Step2: traversal server from 1 to N, if certain j is a server S V _j.memCap > 0 and SV _j.cpuCap > 0, and i ≠ j, then calculate the CPU capacity difference diff of i-th server and a jth server _memwith memory size difference diff _cpu, wherein:

diff _mem＝SV _i.memCap-SV _j.memCap

diff _cpu＝SV _i.cpuCap-SV _j.cpuCap

Step3: for SV _ion each tenant t _ktraversal, if t _kbe Master and t _kdo not belong to SV _j, and meet following condition:

$t_i^k.\mathrm{mem}+{\mathrm{SV}}_i.\mathrm{memCap}>0$ And $t_i^k.\mathrm{cpu}+{\mathrm{SV}}_i.\mathrm{cpuCap}>0$

And ${\mathrm{SV}}_j.\mathrm{memCap}-t_i^k.\mathrm{mem}>0$ And ${\mathrm{SV}}_j.\mathrm{cpuCap}-t_i^k.\mathrm{cpu}>0$

Then calculate and comprise the migration CPU of cost and the mixing calibrator quantity C of internal memory:

$c=[c_{\mathrm{mem}}\&CenterDot;t_i^k.\mathrm{mem}\&CenterDot;({\mathrm{diff}}_{\mathrm{mem}}+t_i^k.\mathrm{mem})+$

$c_{\mathrm{cpu}}\&CenterDot;t_i^k.\mathrm{cpu}\&CenterDot;({\mathrm{diff}}_{\mathrm{cpu}}+t_i^k.\mathrm{cpu})]/t_i^k.\mathrm{dataSize}$

Step4: if by SV _ion certain tenant the c value calculated is minimum in all tenants, then should the object that will move, SV _jit is the destination server that will move.Will and SV _jadd migration strategy S set, and operation below performing:

1. SV _itenant's list in delete

2. SV _jtenant's list in add

3. SV _iinternal memory and CPU capacity deduct internal memory and CPU capacity;

4. SV _jinternal memory and CPU capacity add internal memory and CPU capacity;

5. Step1 is returned.

Step5: if S is not empty, return S; Otherwise return failure.

The invention has the beneficial effects as follows: the virtual machine in physical machine is carried out principal and subordinate's division by the method, virtual machine in each physical machine is made up of from virtual machine (Slave) a host virtual machine (Master) and several, each service operation is on Master, in order to ensure the normal operation of Master due to business when load too high is collapsed suddenly, each Master to have in several Slave distributions in other physical machine, and regularly and Master carry out data syn-chronization.Each physical machine has Node Controller to collect the resource information on each virtual machine, as CPU and internal memory behaviour in service etc.When Node Controller finds that certain physical machine exists focus, then notify cluster manager dual system, cluster manager dual system carries out judging (whether exceeding resource threshold) according to the focus situation in each physical machine, selects suitable physical machine to be moved by focus, thus eliminates focus.The method can eliminate the focus of cluster internal fast, reduces migration cost, realizes load balancing rapidly in server cluster inside.

Below in conjunction with drawings and Examples, the present invention is elaborated.

Accompanying drawing explanation

Fig. 1 is resource composition figure in the hotspot migration method of facing cloud scheduling of resource of the present invention cloud system used.

Embodiment

With reference to Fig. 1.The concrete steps of the hotspot migration method of facing cloud scheduling of resource of the present invention are as follows:

Step one, virtual machine define.

For the server of one group of specified configuration, each server has some virtual machines being called tenant, several tenants form one group, distribute a keeper.

Step 2, default.

5) system comprises N platform physical server, SV ₁... SV _n;

6) each tenant there is memory usage with the utilization rate of CPU the CPU use amount of each server and internal memory use amount are the summations of the CPU of all tenants and the use amount of internal memory.That is: ${\mathrm{SV}}_i.\mathrm{mem}=\underset{k=1}{\overset{\mathrm{Ki}}{\mathrm{\&Sigma;}}}{t}_i^k.\mathrm{mem},$ ${\mathrm{SV}}_i.\mathrm{cpu}=\underset{k=1}{\overset{\mathrm{Ki}}{\mathrm{\&Sigma;}}}{t}_i^k.\mathrm{cpu}.$

SV ₁.cpuCap=80-50-20-20=-10<0

SV ₂.cpuCap=80-20-10-20=30>0

SV ₃.cpuCap=80-30-15=35>0

SV ₄.cpuCap=80-15=65>0

SV ₅.cpuCap=80-15=65>0

Due to SV ₁.cpuCap<0, so SV ₁it is hot spot server.

7) on server, the unit of CPU and internal memory is immutable, and each server has two indices, a memory threshold SV _i.memTh with one CPU threshold value SV _i.cpuTh, this threshold value can not be exceeded, if any one index exceedes on server, claims this server to be a focus.

8) each server has two attributes, memory size SV _i.memCap with CPU capacity SV _i.cpuCap, if SV _i.memCap < 0 or SV _i.cpuCap < 0, this server just becomes hot spot server.Clearly, ${\mathrm{SV}}_i.\mathrm{memCap}={\mathrm{SV}}_i.\mathrm{memTh}-{\mathrm{\&Sigma;}}_{k=1}^{k_i}{t}_i^k.\mathrm{mem},$ And ${\mathrm{SV}}_i.\mathrm{cpuCap}={\mathrm{SV}}_i.\mathrm{cpuTh}-{\mathrm{\&Sigma;}}_{k=1}^{k_i}{t}_i^k.\mathrm{cpu}.$

Step 3, server and virtual machine attribute define.

For each server S V _i, have as properties:

● SV _i.memCap: server S V _iinternal memory residual capacity.

● SV _i.cpuCap: server S V _icPU residual capacity.

● SV _i.memTh: server S V _imemory threshold, exceed then heating point.

● SV _i.memTh: server S V _icPU threshold value, exceed then heating point.

● SV _i.tenants: server S V _ithe set of upper tenant

For each tenant have and each tenant has as properties:

● the identifier of tenant.

● whether tenant is gerentocratic Boolean.

● the memory usage of tenant.

● the CPU usage of tenant.

● the size of the data volume of tenant.

Step 4, basic assumption.

Suppose that X represents any one resource.For N number of numeral { x ₁... x _n, definition mean value variance ${\mathrm{\&sigma;}}_x^2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(\mathrm{xi}-\mathrm{\&mu;x})}^2.$

Therefore, following theorem is obtained:

Theorem 1: if remain unchanged, work as x ₁=x ₂=...=x _ntime minimum.

Theorem 2: for X={x ₁... x _n, obtain:

In formula, be the yardstick measuring resource capacity balance between server, work as x ₁=x ₂=...=x _n=μ _xtime, reach minimum value 0.

In order to the balance between control CPU and internal memory, introduce variable C _memand C _cpu, and be defined as follows:

internal memory correction parameter on individual server.

cPU correction parameter on individual server.

the composed correction parameter of the CPU on individual server and internal memory.

the migration cost of certain tenant.

Step 5, focus remove algorithm.

Algorithm inputs: one group of server S V in cluster _1.... SV _n, wherein may include hot spot server.

Algorithm exports: the strategy set S={ (t, target) of migration }, wherein each strategy is two tuples, comprises the destination server target of a tenant t that will move and its preparation migration.

Initial value: strategy set S=Null; Counter i=0;

Algorithm flow:

Step1: from server list SV _1.... SV _nin obtain i server, judge its memory size SV _ior CPU capacity SV .memCap _i.cpuCap whether be less than 0, if so, then this server is hot spot server;

Step2: traversal server from 1 to N, if certain j is a server S V _j.memCap > 0 and SV _j.cpuCap > 0, and i ≠ j, then calculate the CPU capacity difference diff of i-th server and a jth server _memwith memory size difference diff _cpu, wherein:

diff _mem＝SV _i.memCap-SV _j.memCap

diff _cpu＝SV _i.cpuCap-SV _j.cpuCap

diff _cpu1-2=-10-30=-40；

diff _cpu1-3=-10-35=-45；

diff _cpu1-4=-10-45=-55；

diff _cpu1-5=-10-45=-55。

Screening destination server:

Judge the cpu load of migration each server rear, prevent new focus from generating.Result of calculation is as follows:

SV ₁.cpuCap=-10+50=40>0

SV ₂.cpuCap=30-50=-20<0

SV ₃.cpuCap=35-50=-15<0

SV ₄.cpuCap=65-50=15>0

SV ₅.cpuCap=65-50=15>0

Due to SV ₄and SV ₅at migration back loading capacity >0, therefore new focus can not be formed after migration, all can as destination server, but due to SV ₄on there is the backup physical machine S of M1 ₁₂, Master and the Slave server violating same business can not be placed on the requirement of physical machine on the same stage, therefore only has SV ₅can as destination server.

Step3: for SV _ion each tenant t _ktraversal, if t _kbe Master and t _kdo not belong to SV _j, and meet following condition:

$t_i^k.\mathrm{mem}+{\mathrm{SV}}_i.\mathrm{memCap}>0$ And $t_i^k.\mathrm{cpu}+{\mathrm{SV}}_i.\mathrm{cpuCap}>0$

And ${\mathrm{SV}}_j.\mathrm{memCap}-t_i^k.\mathrm{mem}>0$ And ${\mathrm{SV}}_j.\mathrm{cpuCap}-t_i^k.\mathrm{cpu}>0$

Then calculate and comprise the migration CPU of cost and the mixing calibrator quantity C of internal memory:

$c=[c_{\mathrm{mem}}\&CenterDot;t_i^k.\mathrm{mem}\&CenterDot;({\mathrm{diff}}_{\mathrm{mem}}+t_i^k.\mathrm{mem})+$

$c_{\mathrm{cpu}}\&CenterDot;t_i^k.\mathrm{cpu}\&CenterDot;({\mathrm{diff}}_{\mathrm{cpu}}+t_i^k.\mathrm{cpu})]/t_i^k.\mathrm{dataSize}$

Step4: if by SV _ion certain tenant the c value calculated is minimum in all tenants, then should the object that will move, SV _jit is the destination server that will move.Will and SV _jadd migration strategy S set, and operation below performing:

6. SV _itenant's list in delete

7. SV _jtenant's list in add

8. SV _iinternal memory and CPU capacity deduct internal memory and CPU capacity;

9. SV _jinternal memory and CPU capacity add internal memory and CPU capacity;

10. Step1 is returned.

Step5: if S is not empty, return S; Otherwise return failure.

Claims (1)

1. a hotspot migration method for facing cloud scheduling of resource, is characterized in that comprising the following steps:

Step one, virtual machine define;

For the server of one group of specified configuration, each server has some virtual machines being called tenant, several tenants form one group, distribute a keeper;

Step 2, default;

1) system comprises N platform physical server, SV ₁... SV _n;

2) each tenant there is memory usage with the utilization rate of CPU the CPU use amount of each server and internal memory use amount are the summations of the CPU of all tenants and the use amount of internal memory; That is: ${\mathrm{SV}}_i.mem=\underset{k=1}{\overset{Ki}{\&Sigma;}}{t}_i^K.mem,{\mathrm{SV}}_i.cpu=\underset{k=1}{\overset{Ki}{\&Sigma;}}{t}_i^K.cpu;$ Wherein, that represent is K tenant on the i-th station server;

3) on server, the unit of CPU and internal memory is immutable, and each server has two indices, a memory threshold SV _i.memTh with one CPU threshold value SV _i.cpuTh, this threshold value can not be exceeded, if any one index exceedes on server, claims this server to be a focus;

4) each server has two attributes, memory size SV _i.memCap with CPU capacity SV _i.cpuCap, if SV _ior SV .memCap<0 _i.cpuCap<0, this server just becomes hot spot server; Clearly, ${\mathrm{SV}}_i.memCap={\mathrm{SV}}_i.memTh-{\mathrm{\&Sigma;}}_{k=1}^{k_i}{t}_i^k.mem,$ And ${\mathrm{SV}}_i.cpuCap={\mathrm{SV}}_i.cpuTh-{\&Sigma;}_{k=1}^{k_i}{t}_i^k.cpu;$

Step 3, server and virtual machine attribute define;

For each server S V _i, have as properties:

● SV _i.memCap: server S V _iinternal memory residual capacity;

● SV _i.cpuCap: server S V _icPU residual capacity;

● SV _i.memTh: server S V _imemory threshold, exceed then heating point;

● SV _i.memTh: server S V _icPU threshold value, exceed then heating point;

● SV _i.tenants: server S V _ithe set of upper tenant

For each tenant have wherein, SV _i.tenants represent tenants all on the i-th station server, and each tenant has as properties:

● the identifier of tenant;

● whether tenant is gerentocratic Boolean;

● the memory usage of tenant;

● the CPU usage of tenant;

● the size of the data volume of tenant;

Step 4, basic assumption;

Suppose that X represents any one resource; For N number of numeral { x ₁... x _n, definition mean value variance ${\mathrm{\&sigma;}}_x^2=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}{(xi-\mathrm{\&mu;}x)}^2;$

Therefore, following theorem is obtained:

Theorem 1: if remain unchanged, work as x ₁=x ₂=...=x _ntime minimum;

Theorem 2: for X={x ₁... x _n, obtain:

In formula, be the yardstick measuring resource capacity balance between server, work as x ₁=x ₂=...=x _n=μ _xtime, reach minimum value 0;

In order to the balance between control CPU and internal memory, introduce variable C _memand C _cpu, and be defined as follows:

internal memory correction parameter on individual server;

cPU correction parameter on individual server;

the composed correction parameter of the CPU on individual server and internal memory;

the migration cost of certain tenant; σ ²represent the composed correction parameter of individual server, σ ²t () represents the composed correction parameter of the server at tenant t place;

Step 5, focus remove algorithm;

Algorithm inputs: one group of server S V in cluster _1.... SV _n, wherein may include hot spot server;

Algorithm exports: the strategy set S={ (t, target) of migration }, wherein each strategy is two tuples, comprises the destination server target of a tenant t that will move and its preparation migration;

Initial value: strategy set S=Null; Counter i=0;

Algorithm flow:

Step1: from server list SV _1.... SV _nin obtain i server, judge its memory size SV _ior CPU capacity SV .memCap _i.cpuCap whether be less than 0, if so, then this server is hot spot server;

Step2: traversal server from 1 to N, if certain j is a server S V _jand SV .memCap>0 _j.cpuCap>0, and i ≠ j, then calculate the CPU capacity difference diff of i-th server and a jth server _memwith memory size difference diff _cpu, wherein:

diff _mem＝SV _i.memCap-SV _j.memCap

diff _cpu＝SV _i.cpuCap-SV _j.cpuCap

Step3: for SV _ion each tenant t _ktraversal, if t _kbe Master and t _kdo not belong to SV _j, and meet following condition:

$\begin{array}{ccc}{t}_i^k.mem+{\mathrm{SV}}_i.memCap>0& And& t_i^k.cpu+{\mathrm{SV}}_i.cpuCap>0\end{array}$

$\begin{array}{cccc}And& {\mathrm{SV}}_j.memCap-t_i^k.mem>0& And& {\mathrm{SV}}_j.cpuCap-t_i^k.cpu>0\end{array}$

Then calculate and comprise the migration CPU of cost and the mixing calibrator quantity C of internal memory:

$\begin{array}{c}c=\&lsqb;c_{mem}\&CenterDot;t_i^k.mem\&CenterDot;\left({\mathrm{diff}}_{mem}+t_i^k.mem\right)+\\ c_{cpu}\&CenterDot;t_i^k.cpu\&CenterDot;\left({\mathrm{diff}}_{cpu}+t_i^k.cpu\right)\&rsqb;/t_i^k.dataSize\end{array}$

Step4: if by SV _ion certain tenant the c value calculated is minimum in all tenants, then should the object that will move, SV _jit is the destination server that will move; Will and SV _jadd migration strategy S set, and operation below performing:

1. SV _itenant's list in delete

2. SV _jtenant's list in add

3. SV _iinternal memory and CPU capacity deduct internal memory and CPU capacity;

4. SV _jinternal memory and CPU capacity add internal memory and CPU capacity;

5. Step1 is returned;

Step5: if S is not empty, return S; Otherwise return failure.