CN103257896B - A kind of Max-D job scheduling method under cloud environment - Google Patents

Abstract

The invention discloses the Max D job scheduling method under a kind of cloud environment.This algorithm, by estimating each task operation time in resource, is that the most suitable resource of task choosing performs with Max D-algorithm, and when actual operating efficiency is higher than estimated efficiency, mates task with resource.This job scheduling method can the load balancing of resource under effective guarantee cloud environment, and operation average operating time can be made to reduce, increase the handling capacity of system.

Description

A kind of Max-D job scheduling method under cloud environment

Technical field

The present invention relates to cloud environment job scheduling method, the Max-D job scheduling method under a kind of cloud environment.

Technical background

In recent years, cloud computing mode quickly grows, and framework and the method for operation of IT industry are changed the most therewith.Computing power is wanted by cloud computing Asking reduction, the most supermatic management mode seldom needs manual intervention, substantial saving in enterprise procurement and artificial cost.This also makes High-performance computer, high end storage, high-end server markets are gradually tied up by the cluster of low side devices；Traditional data center is with low cost Cloud computing center replace；A large amount of software application are release in cloud platform in the way of service, and even many application and game can be transported in " cloud " OK.

Cloud computing workload to be dealt with and data volume are huge, and the system almost moment is all processing operation and the data of magnanimity the most how Cloud resource is reasonably distributed, operation is dispatched efficiently, enabling meet the use demand of user, allow operation that user submits to The process time is shorter, execution cost is less, and it is the emphasis in cloud computing and difficult point that the load of simultaneity factor maintains the state of a relative equilibrium.Just It is the demand owing to needing farthest to meet QoS of customer (Quality of Service, QOS) under cloud environment, so research is made under cloud environment Industry dispatching method significant.Inappropriate job scheduling strategy can cause the waste of cloud resource, increases execution time and the cost of user job, Even more so that system congestion cannot continue to provide service time serious；And suitable job scheduling can be on the premise of meeting the demand of user, by resource Waste reduces as far as possible, reaches the expection of user and cloud service provider.Therefore, study under a kind of cloud environment that can meet user and enterprise demand Job scheduling method is necessary.

A.Vouk proposes Min-Min job scheduling method in " Cloud Computing-Issues, Research and Implementations ". Min-Min dispatching method estimates that each unscheduled operation minimum completion time obtains minimum completion time set, then by complete for the minimum of All Jobs The one-tenth time compares, and chooses the minimum operation of deadline in set and is dispatched to suitably calculate node.Min-Min method enables to single work The deadline of industry is less, but can produce the unbalanced of load, and the flat near deadline of operation is longer.

Summary of the invention

It is an object of the invention to provide the job scheduling method under a kind of cloud environment so that when calculating node processing operation under cloud environment, can keep load all Weighing apparatus, and reduce the average completion time of operation.

The technical scheme realizing the object of the invention is:

A kind of Max-D job scheduling method under cloud environment, step is as follows:

The first step: determine all calculating resources and the set of idling-resource in cloud environment.

Second step: submit to priority to be ranked into queue by operation operation to be allocated, the new operation submitted to is added into this queue tail.

3rd step: be scheduling the operation after sequence, uses the suitable resource of Max-D method choice to perform.

For the Max-D method of the 3rd step, its step is as follows:

Step 3.1: to all operations to be allocated, calculates operation averaged power spectrum in all calculating resources and runs the time；

Step 3.2: the averaged power spectrum calculating each operation ran between time and its operation time minimum in the calculating resource of single free time Difference Di, and record this calculating resource；

Step 3.3: find the operation that difference Di is maximum in All Jobs, and this Di is designated as D；

Step 3.4: if D >=0, then assign operation and process to the resource of record, this resource is removed from idling-resource set simultaneously；If D ＜ 0, then redefine resource and the idling-resource set of distribution, joins in idling-resource set, then by the resource completing its distribution operation Return step 3.1.

Step 3.5: repeat step 3.2 to step 3.4 until the resource for all application operations is assigned with operation.

Step 3.1 calculates resource averaged power spectrum deadline method as follows:

Assume that cloud environment is by n unallocated operation T={t₁,t₂,...t_nAnd m resource R={r₁,r₂,...r_mComposition, each resource simultaneously can only One operation is processed；Resource number idle in resource is k, is designated as R'={r₁',r₂',...r_k', wherein k < m；Operation ti is in resource rj On estimation run the time be TCirj, then operation ti average operating time in all resources is

The operation ti deadline in resource rj, be the residual completion time of operation and operation ti just performed on rj on rj when completing Between sum.

Assuming in cloud environment, for same class operation, the data volume that the speed that resource processes processes to it is directly proportional.Operation i is in resource r Estimated Time Of Completion is just to run the residual completion time of operation in resource r and operation i performs time sum in resource r:

${\mathrm{TCir}}_j(k+1)={\mathrm{RTCir}}_j\left(k\right)\&times;\frac{1-pro}{pro}+\&lsqb;(1-\mathrm{\&rho;})\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)pro}\&rsqb;\&times;M(k+1),r_j\&Element;R---\left(1\right)$

Wherein, TCir_j(k+1) deadline needed for representing resource rj process operation ti, TCir_jK () represents that previous operation is in resource rj On the prediction deadline；M (k) is the ratio running this operation required time with run unit operation required time；RTCir_jK () represents previous Individual operation actual run time on rj, pro (0 ＜ pro≤1) represents the completed percentage of previous operation, if resource rj is idling-resource, i.e. Previous operation has performed, then pro=1, and above-mentioned formula can be reduced to

${\mathrm{TCir}}_j(k+1)=\&lsqb;(1-\mathrm{\&rho;})\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)}\&rsqb;\&times;M(k+1),r_j\&Element;R^{\&prime;}---\left(2\right)$

Time TCir is performed by the estimation of operation previous in this resource_j(k) and actual execution time RTCir_jK (), uses formula (1) to carry out Estimation obtains the unscheduled operation execution time in certain resource.But, in the stage that system has just started, each resource was also not carried out operation, Then the execution time of resource cannot be estimated by the implementation status of previous operation.Therefore when system just starts, for all resources, order

TCir_j(0)=RTCir_j(0)=0 (3)

First the most pending operation can select the resource being not carried out operation to perform, and after resource has performed first operation, is just made Actual execution time RTCir of industry_j(1), TCir is made_j(1) equal to RTCir_j(1), then the operation time of operation afterwards is estimated according to formula (1) Calculate.

The method calculating difference D in step 3.2 is as follows:

The operation ti minimum operation time on the node of all unallocated work is designated as mUTC_i=min{TCir₁',TCir₂',...,TCir_k', note TCir is met under record_j'=mUTC_iUnallocated operation rj ', and remember BR_i=r_j', then according to formula D_i=AvgTC_i-mUTC_i, Difference Di to operation i.

Compared with prior art, its remarkable advantage: 1, compared to conventional scheduling method, the job scheduling of the present invention only can be by operation for the present invention It is assigned in the resource of free time, it is ensured that the equilibrium of load under cloud environment, does not haves the situation that part resource transships and other resources are idle；

2, compared to conventional scheduling method, the present invention is that operation selects most suitable resource by Max-D method, decreases averagely completing of operation Time, improve the throughput of system.

Accompanying drawing explanation

Accompanying drawing is the flow chart of Max-D method of the present invention.

Detailed description of the invention

The present invention will be further described below in conjunction with the accompanying drawings.

Assume that cloud environment is by n unallocated operation T={t₁,t₂,...t_nAnd m resource R={r₁,r₂,...r_mComposition, each resource simultaneously can only One operation is processed；Resource number idle in resource is k, is designated as R'={r₁',r₂',...r_k', wherein k < m；Operation ti is in resource rj On estimation run the time be TCirj, then operation ti average operating time in all resources isOperation ti is all The minimum operation time on the node of unallocated work is designated as mUTC_i=min{TCir₁',TCir₂',...,TCir_k', record satisfied TCir_j'=mUTC_iUnallocated operation rj ', and remember BR_i=r_j'；

When needing the operation set non-NULL of scheduling, the following operation of execution:

Step 1: All Jobs in operation set T is calculated AvgTC respectively_i；

Step 2: each operation ti is found mUTC_i, and calculate D_i=AvgTC_i-mUTC_i；

Step 3: find operation ti so that D_i=Max{D₁,D₂,...,D_n, if there being multiple operation to meet condition, then arrive according to these operations Order select ti；

Step 4: if D_i>=0, then assignment operation ti is to the process of resource BRi, resource BRi is removed from idling-resource set R ' simultaneously； If D_i＜ 0, then reappraise resource and the idling-resource set of distribution, joins in idling-resource set, so by the resource completing its distribution operation Rear return step (1).

Step 5: repeat step 2 to step 4 until the resource for all application operations is assigned with operation.

The operation ti deadline in resource rj, be the residual completion time of operation and operation ti just performed on rj on rj when completing Between sum.

Assume that, for same class operation, the data volume that the speed that resource processes processes to it is directly proportional herein.Anticipated in resource r of operation i completes Time is just to run the residual completion time of operation in resource r and operation i performs time sum in resource r:

${\mathrm{TCir}}_j(k+1)={\mathrm{RTCir}}_j\left(k\right)\&times;\frac{1-pro}{pro}+\&lsqb;(1-\mathrm{\&rho;})\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)pro}\&rsqb;\&times;M(k+1),r_j\&Element;R$

Wherein, TCir_j(k+1) deadline distributed to by operation ti needed for resource rj processes, TCir are represented_jK () represents that previous operation is in money The prediction deadline on the rj of source；M (k) is the ratio running this operation required time with run unit operation required time；RTCir_jK () represents Previous operation actual run time on rj, pro (0 ＜ pro≤1) represents the completed percentage of previous operation, if resource rj is idling-resource, The most previous operation has performed, then pro=1, and above-mentioned formula can be reduced to

${\mathrm{TCir}}_j(k+1)=\&lsqb;(1-\mathrm{\&rho;})\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)}\&rsqb;\&times;M(k+1),r_j\&Element;R^{\&prime;}$

According to formula, the unscheduled operation time that performs in certain resource can perform the time by the estimation of operation previous in this resource TCir_j(k) and actual execution time RTCir_jK () is estimated.But, in the stage that system has just started, each resource was also not carried out operation, Then the execution time of resource cannot be estimated by the implementation status of previous operation.Therefore when system just starts, for all resources, order

TCir_j(0)=RTCir_j(0)=0

First the most pending operation can select to be not carried out the resource of operation, after resource has performed first operation, has just obtained the reality of operation Execution time RTCir_j(1), TCir is made_j(1) equal to RTCir_j(1), then the operation time of operation afterwards can be estimated according to formula (1).

Claims (3)

1. the Max-D job scheduling method under a cloud environment, it is characterised in that step is as follows:

The first step: determine all calculating resources and the set of idling-resource in cloud environment；

Second step: submit to priority to be ranked into queue by operation operation to be allocated, the new operation submitted to is added into this queue tail；

3rd step: be scheduling the operation after sequence, uses the suitable resource of Max-D method choice to perform；Max-D method, its step Rapid as follows:

Step 3.1: to all operations to be allocated, calculates operation averaged power spectrum in all calculating resources and runs the time；

Step 3.2: the averaged power spectrum calculating each operation ran between time and its operation time minimum in the calculating resource of single free time Difference Di, and record this calculating resource；

Step 3.3: find the operation that difference Di is maximum in All Jobs, and this Di is designated as D；

Step 3.4: if D >=0, then assign operation and process to the resource of record, this resource is removed from idling-resource set simultaneously；If D ＜ 0, then redefine resource and the idling-resource set of distribution, joins in idling-resource set, then by the resource completing its distribution operation Return step 3.1；

Step 3.5: repeat step 3.2 to step 3.4 until the resource for all application operations is assigned with operation.

Max-D job scheduling method under cloud environment the most according to claim 1, it is characterised in that calculate resource in described step 3.1 Averaged power spectrum deadline method is as follows:

Assume that cloud environment is by n unallocated operation T={t₁,t₂,...t_nAnd m resource R={r₁,r₂,...r_mComposition, each resource simultaneously can only One operation is processed；Resource number idle in resource is k, is designated as R'={r₁',r₂',...r_k', wherein k < m；Operation ti is in resource rj On estimation run the time be TCirj, then operation ti average operating time in all resources is

The operation ti deadline in resource rj, be the residual completion time of operation and operation ti just performed on rj on rj when completing Between sum；

Assuming in cloud environment, for same class operation, the data volume that the speed that resource processes processes to it is directly proportional, and operation i is in resource r Estimated Time Of Completion is just to run the residual completion time of operation in resource r and operation i performs time sum in resource r:

${\mathrm{TCir}}_j\left(k+1\right)={\mathrm{RTCir}}_j\left(k\right)\&times;\frac{1-pro}{pro}+\&lsqb;\left(1-\mathrm{\&rho;}\right)\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)pro}\&rsqb;\&times;M\left(k+1\right),r_j\&Element;R---\left(1\right)$

Wherein, TCir_j(k+1) deadline needed for representing resource rj process operation ti, TCir_jK () represents that previous operation is in resource rj On the prediction deadline；M (k) is the ratio running this operation required time with run unit operation required time；RTCir_jK () represents previous Individual operation actual run time on rj, pro (0 ＜ pro≤1) represents the completed percentage of previous operation, if resource rj is idling-resource, i.e. Previous operation has performed, then pro=1, and above-mentioned formula can be reduced to

${\mathrm{TCir}}_j\left(k+1\right)=\&lsqb;\left(1-\mathrm{\&rho;}\right)\frac{{\mathrm{TCir}}_j\left(k\right)}{M\left(k\right)}+\mathrm{\&rho;}\frac{{\mathrm{RTCir}}_j\left(k\right)}{M\left(k\right)}\&rsqb;\&times;M\left(k+1\right),r_j\&Element;R^{\&prime;}---\left(2\right)$

Time TCir is performed by the estimation of operation previous in this resource_j(k) and actual execution time RTCir_jK (), uses formula (1) to carry out Estimation obtains the unscheduled operation execution time in certain resource；

When system just starts, for all resources, order

TCir_j(0)=RTCir_j(0)=0 (3)

First pending operation can select the resource being not carried out operation to perform, and after resource has performed first operation, has just obtained operation Actual execution time RTCir_j(1), TCir is made_j(1) equal to RTCir_j(1), then the operation time of operation afterwards is estimated according to formula (1).

Max-D job scheduling method under cloud environment the most according to claim 1, it is characterised in that calculate difference D in step 3.2 Method is as follows:

The operation ti minimum operation time on the node of all unallocated work is designated as mUTC_i=min{TCir₁',TCir₂',...,TCir_k', note TCir is met under record_j'=mUTC_iUnallocated operation rj ', and remember BR_i=r_j', then according to formula D_i=AvgTC_i-mUTC_i, Difference Di to operation i.