CN105302650A - Dynamic multi-resource equitable distribution method oriented to cloud computing environment - Google Patents

Abstract

The invention discloses a dynamic multi-resource equitable distribution method oriented to the cloud computing environment. The method is based on user resource needs and the user shared resource quantity, the fairness of resource distribution is guaranteed by constructing a linear programming model, a fast polynomial time complexity algorithm, namely, a bisection method is applied, vt corresponding to a user t is looked for, and an optimum allocation scheme (please see the formula in the specification) is solved to solve the algorithm running efficiency problem of the optimum allocation scheme. According to the method, the practical condition of resource distribution in a cloud computing sharing platform can be reflected accurately and really, the fairness and efficiency of multi-user multi-resource distribution under the dynamic state are improved, and the problem that an existing resource distribution algorithm is too long in running time is solved.

Description

A kind of dynamic multiple resource fair allocat method under facing cloud computing environment

Technical field

The present invention relates to field of cloud computer technology, especially relate to a kind of for the dynamic multiple resource fair allocat method in cloud computing shared platform.

Background technology

Cloud computing shared platform, by integrating multiple computational resource, realizes the unified management to resource and scheduling, for applications provides various service.The resource dispatching strategy of cloud shared platform, according to the task resource demand of each virtual machine node, carries out reasonable distribution to resource, makes virtual machine node that enough resources can be had to complete calculation task and ensures that the fairness of resource between user is distributed.Therefore how under user task resource requirement, justice is carried out to resource and effectively distribute meeting, thus ensure the fairness of Resourse Distribute and to improve allocation efficiency of resource be the key issue that cloud computing shared platform needs to solve.

At present, for solving on the fairness problem of Resourse Distribute, main research work concentrates on following two aspects: single resource and multiple resource distributional equity sex chromosome mosaicism.In single Resourse Distribute is fair, mainly concentrate on and seek in maximization allocation strategy (Max-min) research of minimum resource requirements to single resource (as CPU, link bandwidth etc.), ensure that the resource requirement of most users is met, realize fairness and distribute.In multiple resource fairness in distribution, the people such as Ghodsi systematically study multiple resource fair allocat problem in cloud computing system the earliest, propose multi-user's multiple resource DRF (DominantResourceFairness) algorithm, solve at multiple different resource and how to ensure fairness in distribution sex chromosome mosaicism under depositing situation, demonstrate many gratifying fair properties, and derive various multiple resource fair allocat mechanism (GHODSIA on this basis, ZAHARIAM, HINDMANB, etal.Dominantresourcefairness:Fairallocationofmultiplere sourcetypes [A] .Proceedingsofthe8 ^thuSENIXConferenceonNetworkedSystemsDesignandImplementatio n [C] .Berkeley, CA, 2011.24.).The people such as Dolev propose a kind of (Bottle-basedFairness of the fair allocat based on bottleneck, BBF) algorithm (DOLEVD, FEITELSONDG, HALPERNJY, etal.Nojustifiedcomplaints:Onfairsharingofmultipleresour ces [A] .Proceedingsofthe3rdInnovationsinTheoreticalComputerScie nceConference [C] .Cambridge, MA, 2012.68-75.); The people such as Bhattacharya redefine DRF with this to support multi-level Resourse Distribute (BHATTACHARYAAA, CULLERD, FRIEDMANE, etal.Hierarchicalschedulingfordiversedatacenterworkloads [A] .Proceedingsofthe4thAnnualSymposiumonCloudComputing (SOCC ' 13) [C] .Erlangen, Germany, 2013.); The people such as Wang propose to improve DRF strategy, isomery cloud computing environment (WANGW can be applicable to, LIB, LIANGB.Dominantresourcefairnessincloudcomputingsystemswi thheterogeneousservers [C] .ProceedingsofIEEEINFOCOM [C] .Toronto, Canada, 2014.583-591.); The people such as Psomans study on multiple computing node the multiple resource fair allocat method (PSOMASCA of discrete tasks, SCHWARTZJ.Beyondbeyonddominantresourcefairness:Indivisib leresourceallocationinclusters [R] .TechReportBerkeley, 2013.); The people such as Parks expand DRF by multiple method, comprise the situation (PARKESDC such as the zero demand of some resource, not Divisible loads and weighted user inclination, PROCACCIAAD, SHAHN.Beyonddominantresourcefairness:extensions, limitations, andindivisibilities [J] .ACMTransactionsonEconomicsandComputation, 2015,3 (1): 3.); But these researchs are all static allocation based on user resources demands all in system, to add at random from user in real system and the dynamic that exits has the different of essence, and do not consider history assignment information in resource allocation process.For this problem, researcher proposes resource fairness allocation strategy DDRF under a kind of dynamic case, user is allowed to add system at any time but situation (the KASHI that can not leave, PROCACCIAAD, SHAHN.Noagentleftbehind:Dynamicfairdivisionofmultipleres ources [J] .JournalofArtificialIntelligenceResearch, 2014.579-603.).But researcher brings the hypothesis of equal number resource into when being and adding cloud shared platform based on user, consider the distribution condition of actual user to the many user of shared resource amount in the otherness of resource sharing amount and resource allocation scheduling process, thus result in lower resource utilization and fairness.By finding out the general introduction of above resource fairness allocation strategy, current resource allocation methods does not consider user resources demand and the fairness of shared stock number otherness on Resourse Distribute and the impact of efficiency under dynamic case, therefore be necessary under cloud computing environment, dynamic multiple resource fair allocat method made further research and makes its more realistic resource allocation scheduling demand.

Summary of the invention

In order to solve the problem, the present invention is based on the otherness of user resources demand and shared resource amount under dynamic case, a kind of dynamic multiple resource fair allocat method under facing cloud computing environment is provided, is intended to solve multiple resource distributional equity and efficiency in cloud computing shared platform.

Dynamic multiple resource fair allocat method under facing cloud computing environment of the present invention is built by following step:

Dispose cloud computing experiment porch; Multiple resource distributive property is described and definition; Multi-user's multiple resource Dynamic Fairness Distribution Optimization Model under research cloud computing environment; Design realizes a kind of fast resource allocation algorithm; Carry out emulation platform structure; Test experiments is carried out to performance.

In cloud computing shared platform, user task request is responded by the virtual machine with different resource configuration.M kind hardware resource is comprised, as CPU, internal memory, disk, the network bandwidth etc. in hypothetical resource pond.Make R={1,2 ... m} represents the kind set of resource, U={1,2 ... n} represents that user gathers.The resource requirement vector of user i is D _i=(D _i1..., D _im), wherein D _ijrepresent that the demand of task to resource j of user i accounts for the ratio of the total amount of resource j in whole resource pool, and D _ij> 0.To D _imake normalized, obtain d _i=(d _i1..., d _ij..., d _im), wherein the maximum resource of user i mission requirements amount is superior resources j, i.e. d _ij=1.Make w _irepresent that user i is to the shared resource amount of often kind of resource in resource pool, and when there being k user in system, k ∈ 1 ..., n}, represent that user i distributes the superior resources quota obtained, i.e. executable number of tasks.According to superior resources definition, the superior resources share of user i is:

When there being k user in system, a dynamic multiple resource fair allocat scheme is A ^k, wherein for the Resourse Distribute vector of user i. represent that user i distributes the resource quantity obtained on resource j, to often kind of resource j ∈ R, satisfy condition:

Further, dynamic multiple resource fair allocat Optimized model is built: in conjunction with the dynamic Pareto optimality in dynamic resource fair allocat process, excitation sharing, dynamically operate without envy property, anti-tactic, distribute irreversible attributive character based on user resources demand and shared resource amount, design a linear optimization model, see formula (1), by maximizing M ^kvalue, tries to achieve and maximizes superior resources distribution share number and ensure the fairness of Resourse Distribute as far as possible.

$\left\{\begin{array}{c}\begin{array}{cc}\max imize& M^k\end{array}\\ x_i^k\≥M^k\·w_i,\∀i\≤k\\ x_i^k\≥x_i^{k-1},\∀i\≤k-1\\ \underset{i=1}{\overset{k}{\Σ}}{x}_i^k\·d_{ij}\≤\underset{i=1}{\overset{k}{\Σ}}{w}_i,\∀j\end{array}\right.---\left(1\right)$

Based on this linear programming model, introduce one polynomial time complexity algorithm faster, improve the algorithm operational efficiency problem solving optimal distributing scheme; This algorithm adopts dichotomy thought to find v corresponding to a user τ _τ(v _τcorresponding one value), if made set up, then otherwise according to the first two constraint condition in linear programming model (1), release thus determine and finally obtain

The present invention is directed to multi-user's multiple resource fair allocat problem under cloud computing environment, when system is dynamically entered to user, the distributional equity that the otherness of user resources demand and user institute shared resource amount is brought and efficiency are done to consider, and provide a kind of dynamic multiple resource fair allocat method under facing cloud computing environment.The method provides superior resources under dynamic case and the definition of superior resources share, and builds the linear optimization model that meets fair allocat attribute, distributes in the maximum fairness meeting user resources demand and realize as far as possible superior resources simultaneously.A kind of dynamically multiple resource fair allocation algorithm of design makes Riming time of algorithm complexity be reduced to Ο (n ²logn) (n is total number of users in system).Method provided by the invention can solve the fair allocat problem of multi-user's multiple resource under actual conditions, and improves allocation efficiency of resource.

The present invention has following technological merit:

One: due to the diversity of shared resource type and user resources demand in cloud computing shared system and the otherness of shared resource amount, and user dynamically adds or logs off, the distribution of multi-kind resource is made to be difficult to accomplish Complete fairness, based on this, propose a kind of for the superior resources fair allocat method under dynamic case under depositing situation at multiple different resource, multiple resource problem is transformed into the assignment problem of single resource;

Its two: based on user resources demand and shared resource amount, set up a dynamic multiple resource fair allocat Optimized model, this model meets the dynamic Pareto optimality of fair allocat, excitation is shared, dynamically without envying property, preventing tactic from operating, and in dynamic allocation procedure, distribute irreversible attribute, under meeting true cloud computing environment, user's shared resource also can obtain fair actual conditions of effectively distributing;

Its three: the invention provides a kind of dynamic multiple resource fair allocation algorithm fast, this algorithm, according to linear programming model, adopts dichotomy thought to solve optimal resource allocation scheme, reduces Riming time of algorithm, Algorithms T-cbmplexity is reduced to (n is total number of users in system).

Accompanying drawing explanation

Fig. 1 is the structure schematic flow sheet of the dynamic multiple resource fair allocat method under facing cloud computing environment of the present invention;

Fig. 2 is the schematic flow sheet of dynamic multiple resource fair allocation algorithm fast provided by the invention;

Fig. 3 is that the superior resources quota that obtains of the embodiment of the present invention is based on w _ivariation diagram, w in figure _ifor user i is to the shared resource amount of often kind of resource in resource pool, represent that user i distributes the superior resources quota obtained;

Fig. 4 is 100 user's superior resources quotas in the embodiment of the present invention with the comparative graph of optimal value, in figure, i is user, represent that user i distributes the superior resources quota obtained.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with accompanying drawing and embodiment, the present invention is described in further detail, but scope is not limited to described content.

The whole concept of the dynamic multiple resource fair allocat method under this facing cloud computing environment is: for user task resource requirement under dynamic case and user's shared resource amount, build the dynamic multiple resource fair allocat Optimized model that meets fair allocat attribute, maximization can be executed the task number, and provides a polynomial time complexity algorithm for solving optimal distributing scheme; Meeting task resource demand simultaneously, guarantee fairness as far as possible distribute under maximize superior resources distribution share, and the dynamic multiple resource allocation algorithm provided can improve allocation efficiency of resource, makes Riming time of algorithm can at Ο (n ²logn) realize in.

As shown in Figure 1, the dynamic multiple resource fair allocat method under this facing cloud computing environment is built by following step: dispose cloud computing experiment porch; Multiple resource distributive property is described and definition; Multi-user's multiple resource Dynamic Fairness Distribution Optimization Model under research cloud computing environment; Design realizes a kind of fast resource allocation algorithm; Carry out emulation platform structure; Test experiments is carried out to performance.

One as the present embodiment solves scheme, is described and is defined as multiple resource distributive property:

Under dynamic case, assuming that user is in not order arrival in the same time, namely user k arrives after user k-1, and the resource requirement d of user (i > k) _{> k}unknown; Can not leave after user enters system, and each user task resource requirement amount can not change along with the time; The resource requirement vector of user i is D _i=(D _i1..., D _im), wherein D _ijrepresent that the demand of task to resource j of user i accounts for the ratio of the total amount of resource j in whole resource pool, and D _ij> 0.To D _imake normalized, obtain d _i=(d _i1..., d _ij..., d _im), wherein the maximum resource of user i mission requirements amount is user superior resources j, i.e. d _ij=1.The shared resource amount of user i to often kind of resource in resource pool is expressed as w _i, and order represent that user i distributes the superior resources quota obtained, i.e. executable number of tasks, k is number of users in system, k ∈ 1 ..., n}.The superior resources share of definition user i is: when there being k user in system, obtain a dynamic multiple resource fair allocat option A ^k, $A^k=(A_1^k,...,A_i^k,...,A_k^k),$ Wherein $A_i^k=(h_{i1}^k,...,h_{ij}^k,...,A_{im}^k)$ For the Resourse Distribute vector of user i. represent that user i distributes the resource quantity obtained on resource j, to often kind of resource j ∈ R, satisfy condition:

One as the present embodiment solves scheme, builds a linear programming model method maximizing superior resources quota to be:

Set up a linear programming model such as formula shown in (1), maximize M ^kvalue, finally makes when there is k user in system, maximizes superior resources quota and ensure the fairness of Resourse Distribute as far as possible.

$\left\{\begin{array}{c}\begin{array}{cc}Maximize& M^k\end{array}\\ x_i^k\≥M^k\·w_i,\∀i\≤k\\ x_i^k\≥x_i^{k-1},\∀i\≤k-1\\ \underset{i=1}{\overset{k}{\Σ}}{x}_i^k\·d_{ij}\≤\underset{i=1}{\overset{k}{\Σ}}{w}_i,\∀j\end{array}\right.---\left(1\right)$

Wherein first constraint condition ensures superior resources fair allocat as far as possible, and meanwhile, it is more that the user that shared resource amount is more distributes the superior resources quota obtained.In addition, (1) is given in other two constraint conditions in multiple resource assigning process, namely when there being k user in system, distributing the superior resources quota obtained be not less than in system distribution condition when having k-1 user to any user; When there being k user in system, the stock number that system can be distributed mostly is most this linear optimization model meets dynamic resource fair allocat attribute: dynamically Pareto optimality, excitation share, dynamically without envying the operation of property, anti-tactic, distribute irreversible.

Further, according to linear programming optimum target M ^k, introduce a kind of dynamically multiple resource fair allocation algorithm, realize resource fast allocation.

As shown in Figure 2 the resource allocation process of the present embodiment is further described:

This algorithm is based on maximization M ^ktarget, and constraint condition, find the v that a user τ is corresponding _τif made set up, then $x_i^k=x_i^{k-1};$ Otherwise $x_i^k=M^k\·w_i,$ Thus determine and finally obtain

Step 1: when there being k user in system, to k-1 value carries out non-increasing sequence (i < k-1), i.e. v ₁>=...>=v _k-1, and input user i to each resource j requirement vector d _i;

Step 2: calculate the demand d of user i in often kind of resource _ijthe Resourse Distribute amount total with first superior resources quota product gained, namely

Step 3: judge whether total Resourse Distribute amount exceedes user's shared resource amount in system, namely to often kind of resource j, if set up, forward step 6 to;

Step 4: if be false, the superior resources quota after sequence is interval, namely to any i, i ∈ U, at interval [v ₁, ∞), [v ₂, v ₁] ..., (0, v _k-1], utilize binary search v _τthe interval [LB, UB] at place;

Step 5: judge UB-LB > 1, if set up, jumps to step 4, otherwise end loop;

Step 6: determine M according to formula (2) ^kvalue;

$M^k=\underset{j}{min}\frac{\underset{i=1}{\overset{k}{\&Sigma;}}{w}_i-\underset{i:x_i^{k-1}>v_{\mathrm{\&tau;}}}{\&Sigma;}{x}_i^{k-1}\&CenterDot;d_{ij}}{\underset{i:x_i^{k-1}\&le;v_{\mathrm{\&tau;}}}{\&Sigma;}{w}_i\&CenterDot;d_{ij}}---\left(2\right);$

Step 7: pass through calculate the superior resources quota that each user i is assigned to

Step 8: calculate export the allocative decision of user i on resource j

In the present embodiment Stochastic choice 100 Google clusters, in true calculation task load, required CPU and Memory resource is as user resources request, and dynamic sequential enters system, and arranges the shared resource amount w of each user at random _i, meet dynamic multiple resource fair allocation algorithm is utilized to realize the fast allocation procedures of resource, and the resource allocation conditions of the different sharing stock number that user brings under testing dynamic case.

Fig. 3 is presented at the impact that user's shared resource amount under dynamic case is distributed user resources, evenly have chosen wherein 20 users and distributes the superior resources quota obtained, and to w _icarry out ascending order arrangement.As can be seen from Figure 3, user's shared resource amount w _imore, it is more that it distributes the superior resources quota obtained, and dynamic multiple resource fair allocat method ensure that " fairness " of Resourse Distribute.

Fig. 4 shows 100 users under dynamic multiple resource fair allocation algorithm, is assigned with the superior resources quota the obtained number of tasks of also namely distributing execution and the comparative result separating the optimal value that linear programming system of equations formula (1) obtains; Dynamically multiple resource fair allocation algorithm curve, all the time close to optimal value curve, has taken into account the Efficiency and fairness of Resourse Distribute as can be seen from Figure 4.

Under the experimental result of embodiment is presented at dynamic case, user resources demand can be taken into account by dynamic multiple resource fair allocat method, in turn ensure that the user that shared resource amount is more in fair allocat situation as far as possible can obtain more multiple resource.

From above-mentioned, the embodiment of the present invention is based on cloud shared platform resource allocation system, consider the fairness of otherness on Resourse Distribute and the impact of efficiency of user task resource requirement and the user resources amount of sharing, utilize optimization programming thinking to solve resource fairness assignment problem under dynamic case.

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.All any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included in protection scope of the present invention.

Claims (1)

1. the dynamic multiple resource fair allocat method under facing cloud computing environment, is characterized in that comprising the steps:

(1) based on user resources demand under dynamic case and shared resource amount, build dynamic multiple resource fair allocat linear optimization model, this linear programming model is: $\left\{\begin{array}{c}Maximize{M}^k\\ x_i^k\&GreaterEqual;M^k\&CenterDot;w_i,\&ForAll;i\&le;k\\ x_i^k\&GreaterEqual;x_i^{k-1},\&ForAll;i\&le;k-1\\ \underset{i=1}{\overset{k}{\&Sigma;}}{x}_i^k\&CenterDot;d_{ij}\&le;\underset{i=1}{\overset{k}{\&Sigma;}}{w}_i,\&ForAll;j\end{array}\right.,$

In this linear programming model, k is number of users in system, represent that user i distributes the superior resources quota obtained, w _irepresent that user i is to the shared resource amount of often kind of resource in resource pool, d _ijrepresent that user i task is to the demand of resource j; This linear programming model is determined when there is k user in system, maximizes M ^kvalue, thus determine that each user distributes the superior resources quota obtained

(2) based on above-mentioned linear programming model, optimal distributing scheme is solved by following step

Step 1: when there being k user in system, to k-1 value carries out non-increasing sequence (i < k-1), i.e. v ₁>=...>=v _k-1, and input user i to each resource j requirement vector d _i=(d _i1..., d _ij..., d _im);

Step 2: calculate user i demand d in often kind of resource _ijwith first superior resources quota product gained total resources sendout, namely

Step 3: judge whether total resources sendout exceedes user's shared resource amount in system, namely to often kind of resource j, if $\underset{i=1}{\overset{k}{\&Sigma;}}{v}_1\&CenterDot;d_{ij}\&le;\underset{i=1}{\overset{k}{\&Sigma;}}{w}_i$ Set up, forward step 6 to;

Step 4: if be false, the superior resources quota after sequence is interval, namely to any i, i ∈ U, at interval [v ₁, ∞), [v ₂, v ₁] ..., (0, v _k-1], utilize binary search v _τthe interval [LB, UB] at place;

Step 5: judge UB-LB > 1, if set up, jumps to step 4, otherwise end loop;

Step 6: determine M according to following formula ^kvalue,

Step 7: pass through calculate the superior resources quota that each user i is assigned to

Step 8: calculate export the allocative decision of user i on resource j