CN100576177C - Bidirectional grade gridding resource scheduling method based on the QoS constraint - Google Patents

Abstract

The present invention relates to a kind of bidirectional grade gridding resource scheduling method based on the QoS constraint, its dispatching method is: 1, the QoS request of submitting to according to task is tested the QoS service that all available machines used provide, result with the test of matrix form record, can carry out then and write down the execution time, can not carry out and do respective identification; 2, add up by horizontal direction and vertical direction to of the principle distortion of test result matrix, and to the matrix after the distortion, calculate two vectors by appointment; 3, by two vectors that calculate gained task and resource are carried out non-descending grouping; 4, carry out priority scheduling by the task group result, divide the little priority of class value high scheduling earlier, packets inner is dispatched by Min-min algorithm principle; 5, as if going out current task different resource there is the identical minimum deadline in the scheduling, then dispatches, divide the little priority of class value high scheduling earlier by the resource group result; 6, repeat the 4th step to the five steps, finish up to all task schedulings.

Description

Bidirectional grade gridding resource scheduling method based on the QoS constraint

Technical field

The invention belongs to a kind of grid resource scheduling method, particularly a kind of bidirectional grade gridding resource scheduling method based on the QoS constraint.

Background technology

Grid computing is exactly that resource that distribute on the geography, isomery is linked together, and forms a high performance supercomputer, is available anywhere and reliable computing power for the user provides.Grid provides shared network resource and collaborative platform of dealing with problems, because grid is based upon under the isomerous environment, resource on the network is that different tissues owns, different separately way to manages is arranged, therefore cross-domain resource management has challenge under the grid environment, how to shine upon a group task and be proved to be np complete problem to one group of resource under grid environment.The research of Grid Resource Schedule Algorithms is an important problem of grid resource, has proposed a lot of dispatching algorithms at present both at home and abroad, and existing grid resource scheduling method can simply be divided into two classes: static scheduling method and dynamic dispatching method.

The static scheduling method is the dispatching method that early occurs, and it is comparatively simple, and the operation expense is little, and data dependency is little, thereby is the method that is studied the earliest in grid computing.Common static scheduling method comprises at present: kinds more than ten such as OLB, MET, MCT, Min-Min, Max-Min, Duplex, GA, SA, GSA, Tabu, A*.In grid computing environment,, make the static scheduling method be difficult to design and realization because each processor processing speed difference has been brought for design static scheduling algorithm and is difficult to balanced loaded difficulty.Static scheduling strategy on the heterogeneous platform has run into a lot of problems: the dispatching method of devise optimum is proved to be np complete problem; Be difficult to accurately estimate task execution time and communication delay; Can't tackle the diversity of processor speed.By contrast, dynamic dispatching method has a lot of advantages.

Dynamic dispatching method can solve problems such as load evaluation, effect mensuration, operation transmission, vector calculating, task choosing and task immigration effectively.Dynamic dispatching method can be divided into line model and batch mode, line model is meant as long as task arrives and just is mapped to machine, this pattern is only considered once the mapping of each task, just can not change again in case promptly task is mapped, the heuristic dispatching method of common line model has: OLB, MCT, MET, SA and KPB etc., line model is simpler, but overall performance is not high.Batch mode is meant that first task arrives in the grid and is not mapped to machine immediately, forms a set of tasks but task collected, and just the task in this set concentrated mapping after waiting the mapping event comes.Under the batch mode, dispatching method has obtained the solicited message of task and the actual execution time of considerable task before carrying out, so can make the better scheduling decision-making.Common batch mode dispatching method has: Min-min, Max-min and Sufferage etc., and wherein, Min-min method thinking is simple, and stable performance all has good performance under most environment, be one of research basis of present gridding scheduling algorithm.

Existing gridding scheduling method all is to be developed by the Distributed Calculation classic algorithm, though preferable performance is arranged in distributed environment, but the grid heterogeneous environment for non-centralized management all has limitation: at first, well do not solve for task deadline minimization problem; Secondly, the load balance of isomery machine also can not get guaranteeing in the grid environment; In addition, for the grid resource scheduling problem that the QoS constraint is arranged, said method all can not well solve, and has caused irrational scheduling.Proposed improvement project at these not enough Many researchers of existing methods, mainly be divided into following a few class: to improve the task deadline is the algorithm of purpose; The method that original method is added the QoS constraint; Improve the method for load balance ability; Consider the method for scheduling problem etc. according to economic principle.These methods are having performance preferably aspect a certain specific gridding scheduling problem of solution, but combination property is not strong, is not suitable for the grid environment of reality.Therefore, seek a kind of optimal algorithm of grid environment that is applicable to and be still one of important topic of present grid computing area research.

Summary of the invention

The purpose of this invention is to provide a kind of scheduling that under isomerous environment, realizes the multiple resource under the multi-QoS constraint, shorten the operation deadline, strengthen the bidirectional grade gridding resource scheduling method based on the QoS constraint of the load balance ability of grid resource scheduling.

To achieve these goals, the present invention makes earlier as giving a definition before describing concrete grammar:

Independent task to task definition in the grid environment: a m isomery is expressed as T={t ₀, t ₁..., t _m, setting submitting under the grid environment of task is first task, satisfies following 3 conditions:

(1) each task all be atom and independently, do not have communication and data to rely between the task;

(2) each machine is all monopolized, and promptly when a Task Distribution was given a machine, this task was occupied this machine and finished up to operation;

(3) all be in advance as can be known in the expectation of task before the allocating task on each machine static working time working time promptly.

Definition to resource in the grid environment: the resource in the grid is specially a certain machine, is called for short machine, its expression can provide certain computing power, network storage ability, exact instrument and the resource that some are special.N isomery machine arranged in the hypothetical trellis isomerous environment, and they are expressed as M={m ₀, m ₁..., m _n, the interior at one time task can only carrying out in the grid environment of the machine in the grid is finished the task of could carry out other up to task, and the task of being is monopolized.

Definition to parameter:

Three matrixes and two vectors that two-way graded dispatching method is used are defined as follows respectively:

Definition 1 expection execution time EET (Expected Execution Time) matrix: each element EET _IjExpression resource m _jDo not having under the situation of load, t executes the task _iThe needed time then is the maximal value of system definition if task can not be carried out on machine, and EET is the matrix of a m * n, and it is tested and record by grid management system before scheduling.

Definition 2 expection deadline ECT (Expected Completion Time) matrixes: each element ECT _IjExpression resource m _jT finishes the work _iTime (at t _iBe assigned to m before _jAll Jobs all be performed finish), it also is the matrix of a m * n.

Define the Matrix C carried out oE (the Capability of Execution) matrix of 3 machines: draw each Elements C oE by the EET matrix computations _IjBut the expression resource is to the implementation status of task, and it also is the matrix of a m * n.

To define two vectors in addition: the vectorial NoM of the executable machine quantity of each task (Number of Machines), its each component value is for carrying out the machine quantity of this task; Can the execute the task matrix N oT (Number of Tasks) of quantity of each machine, the quantity of executing the task of its each representation in components corresponding machine.

Set operation t _iTime of arrival be a _i, the time that brings into operation is d _j, then can draw by above definition:

ECT _ij＝d _j+EET _ij.............................................(1)

If operation t _iDistributed to resource m _jOperation makes ECT _iRepresent ECT _Ij, so, maximum execution time just equals the difference between maximum task deadline and the scheduling zero-time, promptly begins to carry out a last operation from first operation and is performed the institute's time spent that finishes, and is zero and has if establish zero-time:

Makespan＝Max(ECT _i)，t _i∈T..................................(2)

Makespan is a criterion of heterogeneous computing system, and the fundamental purpose of grid resource scheduling will reduce Makespan exactly.

The present invention at first to task and resource classification respectively, dispatches task after the classification and resource, and its dispatching method is:

First step: according to the QoS request that task is submitted to the QoS service that all available machines used provide is tested, the result with the test of matrix form record can carry out then and write down the execution time, can not carry out and do respective identification;

Second step: the test result matrix is out of shape by the principle of appointment, and the matrix after the distortion is added up by horizontal direction and vertical direction, calculate two vectors:

A: the number of resources vector that can carry out corresponding task;

B: the executed the task number vector of each resource;

Third step: task and resource are carried out non-descending grouping by two vectors that calculate gained;

The 4th step: carry out priority scheduling by the task group result, divide the little priority of class value high scheduling earlier, packets inner is dispatched by Min-min algorithm principle;

The 5th step: if going out current task has the identical minimum deadline to different resource, then dispatch in the scheduling, divide the little priority of class value high scheduling earlier by the resource group result;

The 6th step: repeat the 4th step to the five steps, finish up to all task schedulings.

The present invention has analyzed under the grid heterogeneous environment complicacy of scheduling of resource under the resource particularity and multi-QoS constraint, solved in the isomerous environment scheduling problem of multiple resource under the multi-QoS constraint, respectively resource and task are carried out classification, realized the effective scheduling under the grid environment.This dispatching method is compared with the conventional mesh resource regulating method, following characteristics are arranged: the complicacy of 1, considering heterogeneous resource from the demand of multi-QoS constraint task, be not limited to the scheduling of traditional scheduler method to computational resource, this patent has proposed corresponding dispatching method to multiple scheduling of resource, more is applicable to grid heterogeneous environment; 2, task has been carried out detailed and rational classification, be different from the two-level scheduler pattern of more original dispatching methods, more can react the characteristic of task, the task that priority scheduling has " harshness " to require has been guaranteed overall deadline minimum; 3, proposed the resource thought of classification is carried out priority scheduling to the scarce resource in the grid environment, improved, and can improve the utilization factor of these resources owing to the unnecessary waiting time of waiting for that some scarce resource produces; 4, respectively task and resource are carried out classification in this dispatching method, comprehensive two-way classification results is dispatched gridding resource, with respect to the traditional scheduler method littler task deadline and load-balancing performance is arranged.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Embodiment

The present invention is described in further detail below in conjunction with accompanying drawing.

The grid resource scheduling model that the present invention is based on QoS is the effective way that solves resource scheduling under the isomerous environment, under this model, task is submitted in the mode of describing its QoS demand, resource is to describe its available QoS method of service issue, and scheduling will guarantee to make under the prerequisite of QoS constraint task that the minimum deadline (Makespan) is arranged as far as possible.Had at present a lot of dispatching methods to consider the problem of QoS constraint, but most influence of all only considering one dimension QoS: simply task is divided into two groups, make the task with high qos requirement be one group at first mapped; Task with end qos requirement is to shine upon after one group.This does not obviously meet the reality of grid environment, grid environment is the isomerous environment that multiple resource constitutes, task can be a various forms to the qos requirement of resource, as: special resources such as cpu performance, machine bandwidth, network capabilities, exact instrument, this just requires grid resource scheduling will consider multidimensional QoS constraint.In the grid environment of considering multidimensional QoS constraint, the QoS service that QoS demand that task is submitted to and resource provide all is diversified, and this makes the complexity that the corresponding situation of task and resource becomes in the grid environment.Resource can only satisfy some QoS demand of task greatly, rather than all, they just can not distribute to this task.This just requires to want the satisfy situation of test resource to task before scheduling, formulate rational scheduling scheme according to the harsh degree and the resource performance difference of task QoS demand.

Two-way graded dispatching method is meant the classification respectively on vertical direction and horizontal direction to expection execution time matrix EET, result according to classification dispatches task, provide the matrix and the vectorial production method that use in the scheduling below, introduced their effects in grid resource scheduling.

The production method of matrix that uses in the scheduling and vector: the QoS service that all available machines used provide is tested according to the QoS request that task is submitted to, and with expecting that execution time matrix EET describes the result of test, wherein can carry out the then record expection execution time, can not executive logging be ' X '.Produce the CoE matrix according to the EET matrix, the situation of each element in the EET matrix is analyzed: if element is numeral in the EET matrix, then the relevant position puts ' 1 ' among the CoE; If element is ' X ' in the EET matrix, then the relevant position puts ' 0 ' in the CoE matrix.Provided a setting of EET matrix below, the CoE matrix of deriving is as follows:

If: $\mathrm{EET}=\left[\begin{array}{ccc}8& X& X\\ 2& 3& 5\\ 4& X& 7\end{array}\right],$ Can derive $\mathrm{CoE}=\left[\begin{array}{ccc}1& 0& 0\\ 1& 1& 1\\ 1& 0& 1\end{array}\right]$

Then, produce NoM vector sum NoT vector by the CoE matrix, the method for generation is: each row element of CoE matrix adds up to each component of NoM vector, and the number of component is identical in number of tasks; Each column element of CoE matrix adds up to each component of NoT vector, and the number of component is identical with number of machines, and formula is as follows:

${\mathrm{NoM}}_i=\underset{j=0}{\overset{m-1}{\mathrm{\Σ}}}{\mathrm{CoE}}_{\mathrm{ij}},(0\≤i\≤n-1).........\left(3\right)$

${\mathrm{NoT}}_j=\underset{i=0}{\overset{n-1}{\mathrm{\Σ}}}{\mathrm{CoE}}_{\mathrm{ij}},(0\≤j\≤m-1).........\left(4\right)$

For top hypothesis to EET, corresponding N oM and NoT vector are respectively: NoM={1,3,2}; NoT={3,1,2}, they are respectively the results that CoE matrix level and vertical direction add up.

Utilize NoM vector sum NoT vector respectively EET to be carried out classification by vertical direction and horizontal direction, method is as follows:

Vertical direction classification: the NoM vector components is carried out non-descending sort, there is the component of identical value to be divided into one group, and write down every group component value as every group scheduling rank: if value is ' 0 ', represent that this task can not finish under current hardware environment, will not dispatch; If other value, then the rank of the more little expression grouping of value is high more, and task is dispatched by height on earth by rank, if the packet-priority that is worth for ' 1 ' is the highest, and scheduling at first.

Horizontal direction classification: the NoT vector components is carried out non-descending sort, have the component of identical value to be divided into one group, and write down every group component value as every group scheduling rank: if value is ' 0 ', represent that this machine can not finish any task, be useless machine; If other value, then the rank of the more little expression grouping of value is high more.

Result according to the vertical direction classification according to priority dispatches, and can guarantee that the task of the harshest QoS constraint is at first dispatched, and prevents that the task of low qos requirement from taking the machine of higher QoS service.Packets inner is pressed the Min-Min algorithmic dispatching, obtains the minimum task deadline as possible.If when using the Min-Min algorithmic dispatching, occur having in the same grouping a plurality of machines simultaneously certain task to be had the minimum deadline, then dispatch by the machine priority that writes down in the NoT vector group result.In the process of whole scheduling, whenever finish the scheduling of a task and will upgrade the value of NoT vector and grouping again, can carry out the last look of residue task quantity as machine to guarantee the value of using at every turn.

The pseudo-code of the two-way graded dispatching method of the present invention is described below:

(1) the All Jobs t among for operation collection T _k

(2) all machine m of for _j

(3)ECT _ij＝EET _ij+d _j

(4)end?for

(5)end?for

(6) go out the CoE matrix by the EET matrix computations

(7) by CoE matrix computations outgoing vector NoM and NoT

(8) operation among the task-set T is worth non-descending grouping by the NoM vector components

(9) removing corresponding NoM component of a vector value among the T is zero task

(10) all tasks are mapped among the do unti1 set of tasks T

(11) all ordering groupings among for task-set T

(12) machine with minimum deadline is found in the operation in each grouping of for

(13) find task t with minimum minimum deadline _k

(14) if If has the identical minimum deadline on a plurality of machines

(15) find the machine m that in the NoT vector, has minimum component value ₁

(16)end?if

(17)end?for

(18) allocating task t _kTo machine m with minimum deadline ₁On

(19) deletion task t from set of tasks T _k

(20) corresponding machine m in the NoT vector ₁Component value subtract 1

(21) upgrade d ₁

(22) value of all i is upgraded ECT _I1

(23)end?for

(24)end?do

From above-mentioned scheduling process as can be seen, this dispatching method is dispatched the few machine of executable machine number earlier, and then the many machines of executable machine number are dispatched, to weaken the correlativity of executing the task by the limited generation of isomery machine performance.Wherein, the 1st row is calculated the expected performance time of this operation on each resource to the 5th row to each concentrated operation of operation.The 6th row is the relational matrix the carried out CoE that the ECT matrix is produced machine to eighth row, and the machine that can satisfy task QoS demand is masked as 1 in matrix, otherwise is 0, and is produced the vectorial NoM of the relation carried out and the NoT of task and machine by the CoE matrix.The task that the 9th row eliminating can not be carried out prevents that task scheduling lost efficacy in cyclic process.The 10th row is that the task of having divided into groups is tested with the Min-min algorithm all available machines to the end, find out the machine assignment operation of minimum deadline, wherein the 14th row is that the machine that the identical minimum execution time is arranged is being selected to 16 row, the principle of selecting is according to NoT matrix grouping level allocation, to reduce machine makespan is exerted an influence to closing property to task.

The content that is not described in detail in this instructions belongs to this area professional and technical personnel's known prior art.

Claims (1)

1, a kind of bidirectional grade gridding resource scheduling method based on the QoS constraint, this method at first to task and resource classification respectively, are dispatched task after the classification and resource, and its dispatching method is:

First step: according to the QoS request that task is submitted to the QoS service that all available machines used provide is tested, the result with the test of matrix form record can carry out then and write down the execution time, can not carry out and do respective identification;

Second step: the test result matrix is out of shape by the principle of appointment, and the matrix after the distortion is added up by horizontal direction and vertical direction, calculate two vectors:

A: the number of resources vector that can carry out corresponding task;

B: the executed the task number vector of each resource;

Third step: task and resource are carried out non-descending grouping by two vectors that calculate gained;

The 4th step: carry out priority scheduling by the task group result, divide the little priority of class value high scheduling earlier, packets inner is dispatched by Min-min algorithm principle;

The 5th step: if going out current task has the identical minimum deadline to different resource, then dispatch in the scheduling, divide the little priority of class value high scheduling earlier by the resource group result;

The 6th step: repeat the 4th step to the five steps, finish up to all task schedulings.

Images