CN116126498A

CN116126498A - Workflow scheduling method oriented to reliability constraint in cloud environment

Info

Publication number: CN116126498A
Application number: CN202310016340.2A
Authority: CN
Inventors: 陈星�; 李金亮
Original assignee: Fuzhou University
Current assignee: Fuzhou University
Priority date: 2023-01-06
Filing date: 2023-01-06
Publication date: 2023-05-16

Abstract

The invention relates to a workflow scheduling method facing reliability constraint in a cloud environment, which comprises the following steps: step S1, constructing a problem model for expressing the execution time and the transmission time of a task by using a triangle ambiguity under the constraint of the overall reliability of a workflow; s2, improving a PSO algorithm and constructing a genetic algorithm-based self-adaptive particle swarm algorithm; and step S3, optimizing the fuzzy completion time and the fuzzy execution cost of the workflow by the self-adaptive particle swarm algorithm based on the genetic algorithm under the condition that the overall reliability constraint of the workflow is met, and obtaining an optimal workflow scheduling scheme. According to the invention, under the condition that the cloud environment has the problems of possible performance fluctuation, downtime and the like of the server, the method has better adaptability to workflow scheduling with reliability constraint.

Description

Workflow scheduling method oriented to reliability constraint in cloud environment

Technical Field

The invention relates to the field of cloud computing workflow scheduling, in particular to a workflow scheduling method oriented to reliability constraint in a cloud environment.

Background

Because cloud data centers have rich computing and storage resources, they are often used to provide services to end users and to address the problem of insufficient computing resources for user terminal devices. The cloud computing technology can flexibly provide resources for users according to the demands of the users, and can execute real-world application programs faster than before, which is important for computation-intensive applications such as astronomy, high-energy physics, bioinformatics, seismic science and the like. These complex computationally intensive applications are composed of hundreds to thousands of interdependent tasks, often built as workflow models. In fact, the scheduling problem of the workflow is of great importance, and the quality of the scheduling result directly affects the completion time and execution cost of the workflow model, especially when the workflow is completed within a reasonable time and budget, still remains a serious challenge.

The goal of the workflow scheduling problem is to select the appropriate computing resources for each task of the workflow to complete the workflow while meeting the needs of the user. Most of scheduling researches of the existing workflow are based on optimal execution cost under the constraint of expiration date or shortest completion time under the constraint of budget. Firstly, the research does not consider that the phenomenon that a server may be down in a real environment and the like can cause that workflow cannot be completed according to a set time, secondly, optimization of a single target often cannot well solve the requirement of a user, and when the completion time is optimized, the cost of renting the server is increased often for obtaining shorter completion time; in optimizing the execution costs, servers with low lease prices but slower processing speeds are often pursued for lower costs. This is due to the nature of the server, which is typically more cost-intensive.

Meanwhile, the existing workflow scheduling problem in the cloud environment is mostly based on an assumption that the task execution time of each workflow on a specific type of virtual machine is deterministic and can be accurately calculated in advance. However, since real world servers cannot maintain persistent performance, actual task execution times may fluctuate while affecting their subtasks. Because the server in the real environment cannot always execute the task according to the predetermined state, an efficient scheduling policy is needed, so that a plurality of Qos can be optimized simultaneously in consideration of the possible problems of the performance, downtime, and the like of the server in the real environment.

Disclosure of Invention

In view of the above, the present invention aims to provide a workflow scheduling method for reliability constraint in cloud environment, which aims to solve the above problems.

In order to achieve the above purpose, the invention adopts the following technical scheme:

a workflow scheduling method facing reliability constraint in cloud environment comprises the following steps:

step S1, constructing a problem model for expressing the execution time and the transmission time of a task by using a triangle ambiguity under the constraint of the overall reliability of a workflow;

s2, improving a PSO algorithm and constructing a genetic algorithm-based self-adaptive particle swarm algorithm;

and step S3, optimizing the fuzzy completion time and the fuzzy execution cost of the workflow by the self-adaptive particle swarm algorithm based on the genetic algorithm under the condition that the overall reliability constraint of the workflow is met, and obtaining an optimal workflow scheduling scheme.

Further, the overall reliability constraint of the workflow is as follows:

the workflow is represented in the form of a directed acyclic graph DAG, i.e., G= < T, E, D >;

wherein T represents a set of nodes, t= { T ₁ ,t ₂ ,...,t _n Each node is a task; e represents a set of edges between tasks, E= { E _1,2 ,e _1,3 ,...,e _i,j ' control or data dependency relationship between tasks, task t _i And task t _j Data size e for transmission between _i,j ＝(t _i ,t _j ) A representation; d= { D (t) ₁ ),d(t ₂ ),...,d(t _n ) And represents the computational workload of the task. From the above definition, it is possible to derive task t _i Is a direct precursor task P (t) _i )＝{t _k |e _k,i -a }; each workflow has a set reliability Rel value, and when the reliability of the scheduling strategy meets a given reliability constraint, the scheduling strategy is considered to be feasible;

let the resources of cloud environment consist of m different virtual machine instance types, with r= { R ₁ ,r ₂ ,...,r _m -representation; for resource r _i By using

Representation of->

Representing resource r _i Is (are) open time, < >>

Representing resource r _i Is set to be off time, u _i Representing resource r _i C _i Representing resource r _i Price per unit time, epsilon _i Representing resource r _i The computing power of different examples is different;

let task t _i Deployment to resource r _j On, task t _i The execution time of (2) is:

for task t _i Parent task P (t) _i )＝{t _p |e _p,i Parent task t _p To task t _i Transmission time trans (t) _p ,t _i ) The method comprises the following steps:

wherein beta represents r (t) _p ) And r (t) _i ) Bandwidth between;

considering the data dependency relationship among tasks, namely, the subtasks are started only after the father task is completed completely, task t _i The start time of (2) is defined as follows:

ST(t _i )＝max{max(FT(t _p )+trans(t _p ,t _i )),Ava(r(t _i ))}

Ava(r(t _i ) R (t) represents a virtual machine _i ) Ready to execute task t _i Is the earliest time of FT (t) _p ) Representing task t _p And then task t _i The completion time of (2) is:

FT(t _i )＝ST(t _i )+ET(t _i ,r(t _i ))

thus, the overall execution time of the workflow is:

T _total ＝max{FT(t _i )|t _i ∈T}

the workflow execution cost comprises calculation cost and data transmission cost, and the execution cost is as follows:

wherein ,c_j,k Represented as resource r _j Transmitting 1GB data to r _k Lambda is the required price of (1) _rj For resource r _j When task i and task j are scheduled on different virtual machine instances, s _i,j =1, otherwise s _i,j ＝0。

Considering the task execution failure caused by faults, the instantaneous faults are set to follow poisson distribution and are formed by the resource r _i Failure rate of epsilon _i Task t _i At resource r _j The reliability of the execution is as follows:

from the additivity of poisson distribution, the overall reliability of the workflow is:

based on the above definition, the optimal scheduling problem of completion time and execution cost under the overall reliability constraint of the workflow can be formally expressed as:

wherein Rel is the overall reliability value and sigma of the workflow of the current scheduling scheme _rel Is a predefined reliability constraint threshold.

Further, a fuzzy theory is introduced, and the calculation time and the transmission time of the task are represented by using the triangular fuzzy number

Is u (x), where t ^m The execution time predefined for the task, left and right endpoints t ^l and t^u Representing a range of variation in task execution time

Using

Triangle ambiguity representing scalar τ, based on the concepts of run-time and transfer-time uncertainty, the completion time and execution cost of the workflow are both triangle ambiguities, expressed as +.>

and

The optimization problem herein is formally expressed as:

for optimization targets

The value is a triangular fuzzy number, the value is represented by the mean +.>

Sum of variances->

Determining together; optimization objective->

The calculation mode of (a) is as follows:

for the mean value

Sum of variances->

For the mean and standard deviation of the fuzzy sets under uniform and proportional distribution, respectively, the triangular fuzzy number +.>

Is a situation based on a proportional distribution, so that the mean +.>

Sum of variances->

Calculated from the following formula:

eta is standard deviation

Weights of (2); for triangle blur number ++>

The treatment method is the same as->

Further, for the estimated time t, the corresponding triangle ambiguity number

t ^m For the most probable execution time of a task, i.e. the execution time of a given task on the server, t ^l and t^u The values respectively from the interval [ delta ] ₁ ×t ^m ,t ^m] and [t^m ,δ ₂ ×t ^m ]Inner random selection, wherein delta ₂ ＞1＞δ ₁ The method comprises the steps of carrying out a first treatment on the surface of the The operation of defining the trigonometric function is as follows:

addition of the triangular fuzzy number:

comparison operation of triangle fuzzy numbers: if it is

Then->

The number multiplication operation of the triangle fuzzy number:

further, the genetic algorithm-based adaptive particle swarm algorithm is specifically as follows:

(1) Using a two-dimensional discrete particle coding mode composed of cloud computing resources and tasks, one particle represents one solution in a problem space, and the position of a particle i at the time t is as follows:

wherein ,

the number of the virtual machine where the 1 st task of the i-th particle is located at the time t is represented;

(2) The objective is to optimize the total cost f of the workflow under the overall reliability constraint of the workflow, wherein the total cost f comprises the fuzzy execution cost of the workflow

And fuzzy execution time->

Two scheduling targets belong to the multi-target planning problem, so the fitness function is set as:

wherein k₁ and k₂ Weight coefficients, T, representing completion time and execution cost, respectively _one ,C _one Representing the completion time and execution cost spent by all tasks executing on only one server;

(3) The crossover operator and mutation operator of genetic algorithm are introduced, and the updating mode of the particle i in the t-th iteration is shown as follows, wherein

And +.Operator and mutation operator:

mutation operators are introduced into the inertia part of the traditional PSO updating formula, and the updating mode is as follows:

wherein ,r₁ Is a random number of (0, 1), p _m For a given probability of variation, when r ₁ <p _m Then

The operation of the mutation is carried out,

will randomly change->

If r ₁ ≥p _m No variant behavior occurs;

for the personal cognition part and the social cognition part, a crossover operator is introduced to update the corresponding part of the traditional updating formula, and the updating mode is as follows:

wherein, the two formulas update the personal cognitive part and the social cognitive part respectively, r ₂ and r₃ Is a random number of (0, 1), p _c For a given crossover probability, when r ₂ (or r) ₃ )<p _c In the time-course of which the first and second contact surfaces,then

The mutation operation is carried out, C _p (or C) _g ) 2 bits of the particle are randomly selected, the server code between the bits of the particle is encoded with +.>

(or gBest) ^t-1 ) The server codes between the corresponding split bits are interleaved.

(4) New adjustment strategy of inertia factor w can be based on current particles

Global best particle gBest with history ^t-1 The difference of w is adaptively adjusted, and the updating mode of w is as follows:

wherein ,

indicating particle->

Global best particle gBest with history ^t-1 The number of different sub-bits in between, |T| represents the number of sub-tasks in the workflow;

the updating mode of the personal cognitive factors and the social cognitive factors adopts a linear increase and decrease strategy. The updating mode is as follows:

wherein ,

and

Respectively the parameter c ₁ And parameter c ₂ Initial value of setting, ++>

and

C is ₁ and c₂ Final value of (2).

Further, the specific cases of two particles to be compared are divided into the following three cases:

1) For two particles to be compared, if both particles meet reliability constraints, selecting particles with smaller total cost

2) If one particle satisfies the reliability constraint and one particle does not satisfy the reliability constraint, selecting a particle that satisfies the reliability constraint

3) If neither particle satisfies the reliability constraint, a particle with high reliability is selected for the particles that do not satisfy the constraint, since the particle is more likely to become a viable solution after iteration, then

Further, the step S3 specifically includes:

1) Initializing related parameters of a genetic algorithm-based adaptive particle swarm algorithm, such as population size PN and maximum iteration times ^M ax _iter Inertia factor w, and randomly generating a population;

2) Calculating fitness, wherein the initial state of each particle is an individual optimal particle, and the particle with the smallest fitness value in the initial population is set as a global optimal particle;

3) Introducing variation of a genetic algorithm and a crossover operator to update the position of the particle, and calculating the adaptability of the updated particle;

4) If the fitness of the updated particles is smaller than that of the individual optimal particles, updating the individual optimal particles, and setting the current particles as the individual optimal particles;

5) Meanwhile, comparing the fitness of the updated particles with the fitness of the global optimal particles, if the fitness of the current particles is smaller than the fitness of the global optimal particles, updating the global optimal particles, setting the global optimal particles for the current particles and updating the optimal fitness;

6) Checking whether the algorithm iteration ending condition is met, if so, ending the algorithm, otherwise, returning to the step 3).

Compared with the prior art, the invention has the following beneficial effects:

aiming at the problems that performance fluctuation, downtime and the like possibly occur in the running process of a virtual machine, the invention designs a problem model for expressing the execution time and the transmission time of a task by using a triangle ambiguity representation under the overall reliability constraint of a workflow, and simultaneously provides an APSOGA algorithm to optimize the completion time and the execution cost of the workflow. The algorithm is based on the traditional PSO algorithm, and mutation operation and crossover operation of a genetic algorithm are added to avoid the algorithm from falling into local optimum. The scheduling test of scientific workflows under five different specifications of three scheduling strategies shows that the APSOGA strategy has better adaptability to workflow scheduling with reliability constraint under the problems that a cloud environment has the server that performance fluctuation, downtime and the like are likely to occur.

Drawings

FIG. 1 is a scheduling model framework in accordance with one embodiment of the present invention

FIG. 2 is a diagram of a triangle fuzzy number membership function in an embodiment of the present invention

FIG. 3 is a mapping relationship of encoded particles according to an embodiment of the present invention;

FIG. 4 shows a variation of particles according to an embodiment of the present invention;

FIG. 5 is a personal cognitive portion crossover operation in one embodiment of the invention;

FIG. 6 is a social cognition portion crossover operation in one embodiment of the invention;

fig. 7 is a block diagram of 5 scientific workflows in one embodiment of the invention.

Detailed Description

The invention will be further described with reference to the accompanying drawings and examples.

Referring to fig. 1-7, the present invention provides a workflow scheduling method for reliability constraint in a cloud environment, which includes the following steps:

In this embodiment, the workflow scheduling model framework is mainly composed of three parts: cloud environment resources, workflows with reliability constraints, and schedulers, as shown in fig. 1;

in this embodiment, the workflow is represented in the form of a directed acyclic graph DAG, i.e., G= < T, E, D >. Wherein T represents a set of nodes, t= { T ₁ ,t ₂ ,...,t _n Each node is a task. E represents a set of edges between tasks, E= { E _1,2 ,e _1,3 ,...,e _i,j ' control or data dependency relationship between tasks, task t _i And task t _j Data size e for transmission between _i,j ＝(t _i ,t _j ) And (3) representing. D= { D (t) ₁ ),d(t ₂ ),...,d(t _n ) And represents the computational workload of the task. From the above definition, it is possible to derive task t _i Is a direct precursor task P (t) _i )＝{t _k |e _k,i }. Each workflow will have a set reliability Rel value, which is considered to be viable when the reliability of the scheduling policy meets a given reliability constraint.

Cloud platforms typically provide computing resources to users in the form of virtual machines for which there are several characteristics. (1) Once assigned to a virtual machine for execution, the task will execute entirely on that virtual machine. (2) The virtual machine can only execute one task at a time, and cannot execute a plurality of tasks at the same time. The resources of the cloud environment proposed herein consist of m different virtual machine instance types, with r= { R ₁ ,r ₂ ,...,r _m And } represents. For resource r _i Can be used

Representation of->

Representing resource r _i Is (are) open time, < >>

Representing resource r _i Is set to be off time, u _i Representing resource r _i C _i Representing resource r _i Price per unit time, epsilon _i Representing resource r _i Is different from instance to instance in terms of computing power.

Suppose task t _i Deployment to resource r _j On, task t _i The execution time of (2) is:

wherein beta represents r (t) _p ) And r (t) _i ) Bandwidth between them. Considering the data dependency relationship among tasks, namely, the subtasks can be started only after the father task is completely finished, task t _i The start time of (2) is defined as follows:

ST(t _i )＝max{max(FT(t _p )+trans(t _p ,t _i )),Ava(r(t _i ))}

FT(t _i )＝ST(t _i )+ET(t _i ,r(t _i ))

thus, the overall execution time of the workflow is:

T _total ＝max{FT(t _i )|t _i ∈T}

Whether a workflow can be completed on time according to the needs of a user is a primary concern for the user. Due to factors such as virtual machine crashes and software defects, tasks may not be completed on the virtual machine according to a preset time. In the studies herein, failure to perform tasks due to failure should be considered. In general, transient faults follow a poisson distribution, defined by the resource r _i Failure rate of epsilon _i Task t _i At resource r _j The reliability of the execution is as follows:

from the formula, it can be known that the reliability of a single task is determined by the execution time of the task on the virtual machine and the failure rate of the virtual machine, and if the execution time of the task is longer and the failure rate of the machine is higher, the reliability is lower.

based on the above definition, attention is paid herein to the optimal scheduling problem of completion time and execution cost under the overall reliability constraint of the workflow, formally expressed as:

In this embodiment, in the workflow scheduling problem in the deterministic cloud environment, we always assume that the performance of the server is not affected by external factors, and the execution time and transmission time of the task on the server can be determined in advance. However, the server performance of a real environment may cause the execution time of a task to be unequal to a predetermined time due to factors such as fluctuation.

Therefore, fuzzy theory is introduced herein, and the computation time and the transmission time of a task are represented by using triangular fuzzy numbers. Triangle fuzzy number

The membership function u (x) of (2) is as shown in fig. 2:

wherein ,t^m The execution time predefined for the task, left and right endpoints t ^l and t^u Indicating the range of variation in task execution time.

Unified use

The triangle blur number representing the scalar τ. Based on the concepts of run-time and transfer-time uncertainty, the completion time and execution cost of the workflow are both triangle ambiguities, expressed as +.>

and

The optimization problem herein can be formally expressed as: />

For optimization targets

Sum of variances->

And (5) jointly determining. Palaios et al propose a comparison criterion to minimize the linear combination of two target valuesKnown as palaios criteria. Thus, optimize goal->

The calculation mode of (a) is as follows:

for the mean value

Sum of variances->

Lee et al define the mean and standard deviation of the fuzzy sets under uniform and proportional distributions, respectively, the triangular fuzzy number +.>

Is a situation based on a proportional distribution, so that the mean +.>

Sum of variances->

Calculated from the following formula:

_η is the standard deviation

Is a weight of (2).

For triangle blur number

The treatment method is the same as->

In this embodiment, a more practical blurring method is provided to describe the execution time and transmission time of the task, and for the estimated time t, the corresponding triangle blur number is provided

t ^m For the most probable execution time of a task, i.e. the execution time of a given task on the server, t ^l and t^u The values respectively from the interval [ delta ] ₁ ×t ^m ,t ^m] and [t^m ,δ ₂ ×t ^m ]Inner random selection, wherein delta ₂ ＞1＞δ ₁ 。

In the workflow scheduling, some operation is required to be performed on the triangle fuzzy number, so we redefine some operation on the fuzzy number.

And adding the triangular fuzzy number. For two triangle ambiguities

and

According to the fuzzy theory

The fuzzy number addition principle defined in the above can be used for obtaining the addition rule of the fuzzy number as follows:

and (5) comparing the triangle fuzzy numbers. Inspired by the comparison criterion proposed by the prior study to compare the magnitude of two fuzzy numbers, the comparison principle adopted here is as follows, if

Then->

And (5) performing the number multiplication operation of the triangle fuzzy number. The number multiplication operation of the triangle fuzzy number is given by the formula:

in the present embodiment, the adaptive particle swarm algorithm APSOGA based on the genetic algorithm is specifically described from the following 5 sections:

(1) Question coding

Because of the very large scale of cloud computing resources, the construction of the resource pool can greatly impact the search efficiency of the algorithm. In order to improve the searching efficiency of the algorithm, a reasonable coding mode is needed to be designed so that the algorithm can better solve the discrete optimization problem of workflow scheduling. The two-dimensional discrete particle coding mode composed of cloud computing resources and tasks is used, one particle represents one solution in a problem space, and the position of a particle i at the time t is as follows:

wherein ,

the virtual machine number at which the 1 st task of the i-th particle is located at time t is represented. Fig. 3 illustrates the encoded particles corresponding to one workflow scheduling policy comprising 5 subtasks. Taking task 1 as an example, the virtual machine number corresponding to task 1 is 4, which means that task 1 will be executed on the virtual machine with number 4 allocated to the resource pool.

(2) Fuzzy fitness function

In this embodiment, the objective is to optimize the total cost f of the workflow under the overall reliability constraint of the workflow, where the total cost f includes the fuzzy execution cost of the workflow

And fuzzy execution time->

The two scheduling targets belong to the multi-target planning problem, so the fitness function is set as follows:

wherein k₁ and k₂ Weight coefficients, T, representing completion time and execution cost, respectively _one ,C _one Indicating the completion time and execution cost spent on executing all tasks on only one server. Since there may be particles in the iterative process that do not meet the reliability constraint, it follows that there may be unfeasible solutions in the candidate solutions that do not meet the reliability constraint. Therefore, for the total cost of comparing two particles in an algorithm, it must be considered whether a particle is a viable solution. The specific case of two particles to be compared is divided into the following three cases.

2) If one particle satisfies the reliability constraint and one particle does not satisfy the reliability constraint, selecting a particle that satisfies the reliability constraint.

3) If neither particle satisfies the reliability constraint, a particle with greater reliability is selected herein for a particle that does not satisfy the constraint, since the particle is more likely to become a viable solution after iteration, then

(3) Particle update strategy

When the algorithm searches for the optimal solution, the particles need to iteratively update their own speed and position. The traditional PSO updating self-position mode has the defect of early convergence, in order to avoid the early convergence of the algorithm, APSOGA introduces a crossover operator and a mutation operator of a genetic algorithm, and updates the corresponding part of the same updating formula. At the t-th iteration, the update mode of the particle i is shown as follows, wherein

And ∈h indicates the crossover operator and mutation operator, respectively.

The APSOGA introduces a mutation operator in the inertia part of the traditional PSO updating formula, and the updating mode is as follows:

The operation of the mutation is carried out,

will randomly change->

If r ₁ ≥p _m Does not send outMutation behavior. Fig. 4 is a variation operation on the encoded particles of fig. 3.

wherein, the two formulas update the personal cognitive part and the social cognitive part respectively, r ₂ and r₃ Is a random number of (0, 1), p _c For a given crossover probability, when r ₂ (or r) ₃ )<p _c When in use, then

(or gBest) ^t-1 ) The server codes between the corresponding split bits are interleaved. Fig. 5 is a cross-over operation of the personal cognitive part of fig. 3, and fig. 6 is a cross-over operation of the social cognitive part of fig. 3.

(4) Parameter adjustment

The inertial weight factor w can determine the convergence and search capabilities of the PSO. For the inertia factor w, when w is smaller, the algorithm has stronger local searching capability, and when w is larger, the algorithm has stronger global searching capability. The following formula is an inertia adjustment strategy of the conventional PSO algorithm.

wherein ,w_max and w_min Respectively, the maximum value and the minimum value of w set during initialization, iter _cur and iter_max The current iteration number and the maximum iteration number of the algorithm.

The inertia factor w of the conventional PSO is only related to the number of iterations, and cannot well satisfy the complexity of the practical problem. The embodiment provides a new adjustment strategy of inertia factor w, which can be based on the current particle

Global best particle gBest with history ^t-1 The difference of w is adaptively adjusted to enhance the searching capability of the APSOGA algorithm. The update mode of w is as follows:

wherein ,

indicating particle->

Global best particle gBest with history ^t-1 The number of different quantiles between, |T| represents the number of subtasks in the workflow. When->

When the value of the (b) is smaller, the number of different bits of the current particle and the historical global optimal particle is smaller, so that the value of w should be reduced, the local searching capability of the algorithm is enhanced, the convergence effect of the algorithm is improved, and the optimal solution is found; otherwise, the value of w should be increased, so as to enhance the global searching capability of the algorithm and enlarge the searching space of the algorithm.

In addition, the personal cognition factor and the social cognition factor of the algorithm are updated by adopting a linear increase and decrease strategy. The update method is as follows.

wherein ,

and

and

C is ₁ and c₂ Final value of (2).

(5) Particle to scheduling result mapping

First, for the encoded particle i, a mapping of the encoded particle i to the scheduling result is given, as specifically shown in algorithm 1:

algorithm 1 mapping of encoded particles to scheduling results

Input: (W, R, X)

And (3) outputting: f (F)

In addition, the main flow of the workflow fuzzy scheduling strategy based on the APSOGA comprises the following 6 steps.

1) Initializing related parameters of APSOGA, such as population size PN, maximum iteration number Max _iter Inertia factor w, etc., anda population is randomly generated.

2) And calculating fitness, wherein the initial state of each particle is an individual optimal particle, and the particle with the smallest fitness value in the initial population is set as a global optimal particle.

3) And introducing mutation of a genetic algorithm and a crossover operator to update the position of the particle, and calculating the fitness of the updated particle.

4) And if the fitness of the updated particles is smaller than that of the individual optimal particles, updating the individual optimal particles, and setting the current particles as the individual optimal particles.

5) And comparing the fitness of the updated particles with the fitness of the global optimal particles, if the fitness of the current particles is smaller than the fitness of the global optimal particles, updating the global optimal particles, setting the global optimal particles for the current particles, and updating the optimal fitness.

Example 1:

the workflow test model used in this embodiment employs workflows from 5 different fields of study by Bharathi et al: cyberShake, bioinformatics Sipht, astronomical Montage, gravitational physics LIGO, and biogenic Epigenomics. The five workflows each have their different structures, and the structures of the five workflow samples are given in fig. 7. Details of these workflows are stored in xml format files.

For different workflows, three workflows of different specifications are selected herein: a micro workflow of about 30 tasks, a mini workflow of about 50 tasks, and a medium workflow of about 100 tasks. The Wen Yun resource pool has 6 cloud servers, as shown in table 1, assuming that the computing capability of the server m4.16xlar is strongest, the computing time of each task of the workflow in the m4.16xlar is directly obtained from a corresponding xml file, and the execution time of each task on other servers can be obtained according to the performance ratio of other servers to the m4.16xlar server.

Six virtual machines were selected from Amazon EC2 cloud platformSimulation experiments were performed, and virtual machine resource configuration information is shown in table 1. Amazon EC2 usually takes 60s or 1h as the asking price interval lambda _i Here, 60s is selected as the asking price interval, and the price of transmitting 1GB data is 0.2.

Table 1 virtual machine configuration information table

Table 1 Virtual machine configuration information table

Reliability sigma of workflow _rel Set to the reliability of the workflow running on the server m4. Xlage at a given time, T _one and C_one Completion time and execution cost required for the workflow to run on a server of m4.2 x-range. For converting the execution time and transmission time of a task into a triangle ambiguity number, a parameter delta ₁ And parameter delta ₂ The weight coefficient eta of the standard deviation is set to be 0.85 and 1.2 respectively, and 1 is taken. Fitness function weight coefficient k ₁ and k₂ 0.2 and 0.3, respectively.

In order to evaluate the effectiveness of the APSOGA algorithm, in this embodiment, the APSOGA strategy is compared with a conventional PSO strategy and a random strategy to perform experimental effects. In existing workflow scheduling problems, these algorithms are typically used as comparison algorithms.

In the traditional PSO strategy, the same coding mode as APSOGA is adopted, the traditional updating mode is adopted, the traditional PSO algorithm parameter setting refers to the existing research, the population size is set to be 50, the maximum iteration number is 500, and c ₁ ＝1，c ₂ ＝1，w＝1。

Random strategy: the method is based on a random search strategy, adopts the same coding mode as APSOGA, adopts a random mode to update particle codes, does not affect each other in each iteration, performs random search in a solution space of a problem, calculates the fitness of each particle, and records the optimal solution in the search process.

In order to test the workflow scheduling performance of the apsog strategy, the PSO strategy and the random strategy under the factors of network fluctuation, server faults and the like in the cloud environment, in this embodiment, three strategies are used to respectively perform 30 groups of repeated experiments on five workflows of different scales, and table2, table 3 and table 4 record the optimal fitness value and the average fitness value of the workflow scheduling experiments of the three strategies, wherein the fitness value represents the weighted sum of the completion time and the execution cost.

The scheduling results of the APSOGA strategy, PSO strategy and random strategy on 30 replicates of five micro-workflows are shown in table 2. It can be derived from the table that, for the micro workflow, the apsog strategy can obtain the optimal solution no matter the average fitness value or the optimal fitness value, the PSO strategy is inferior, and the random strategy is worst, because the apsog strategy adds the mutation operation cross operation of the genetic algorithm on the basis of the PSO strategy, the apsog algorithm has better global optimizing capability than the PSO algorithm. While the effect of the random strategy is poor because the solution efficiency of the random strategy in the search problem space is low.

The scheduling result of 30 repetitions of the small workflow is shown in table 3, and the APSOGA strategy obtains an optimal solution at both the optimal fitness value and the average fitness value. Furthermore, the optimum fitness value of the APSOGA strategy is superior to the traditional PSO strategy by up to 4.7%, up to 20% due to the random strategy.

TABLE2 comparison of micro workflow Experimental Effect Table2 Compositionofmicro-workflow-experiment effects

The scheduling results of 30 times of repetition of the middle-sized workflow are shown in table 4, the APSOGA strategy is superior to other algorithms to different degrees, the performances of the random strategy in the middle-sized workflow and the micro-sized workflow are observed, the random strategy is larger with the increase of the number of tasks, the performance is poorer and the difference between the experimental effect and the experimental effect of the APSOGA strategy is larger with the increase of the number of tasks. In summary, the APSOGA strategy has better scheduling performance than other scheduling strategies, both in small-scale workflows and large-scale workflows.

TABLE 3 comparison of Small workflow experiment Effect Table 3 Compositionofsmalll-workflow xypersendenteffects

The APSOGA strategy introduces a mutation operator and a crossover operator of a genetic algorithm on the basis of the traditional PSO strategy, so that the algorithm avoids the problem of premature convergence of the traditional PSO strategy, avoids particles from being trapped into local optimum, and enables the particles to have better global optimizing capability. Secondly, the APSOGA strategy adjusts the updating mode of the inertia factor w, the traditional PSO strategy adopts a linear mode to update the inertia factor w, w is only related to the current iteration times and cannot well solve the complex problem, and the APSOGA strategy adopts a self-adaptive adjusting mode, considers the different numbers of the current particles and the overall optimal particles, and improves the searching performance of w on the algorithm. Experimental results show that the APSOGA strategy can obtain a better scheduling scheme compared with the traditional PSO strategy or the random strategy.

TABLE 4 Medium-sized workflow experiment Effect vs Table 4Comparsion ofmedium-workflow polymerization effects

The foregoing description is only of the preferred embodiments of the invention, and all changes and modifications that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.

Claims

1. A workflow scheduling method facing reliability constraint in cloud environment is characterized by comprising the following steps:

2. The workflow scheduling method for reliability constraint in cloud environment according to claim 1, wherein the overall workflow reliability constraint is as follows:

Representation of->

Representing resource r _i Is (are) open time, < >>

wherein beta represents r (t) _p ) And r (t) _i ) Bandwidth between;

ST(t _i )＝max{max(FT(t _p )+trans(t _p ,t _i )),Ava(r(t _i ))}

FT(t _i )＝ST(t _i )+ET(t _i ,r(t _i ))

thus, the overall execution time of the workflow is:

T _total ＝max{FT(t _i )|t _i ∈T}

3. The workflow scheduling method for reliability constraint in cloud environment as recited in claim 1, wherein fuzzy theory is introduced, and a triangle fuzzy number is used to represent calculation time and transmission time of task, the triangle fuzzy number

Using

and

The optimization problem herein is formally expressed as:

for optimization targets

Sum of variances->

Determining together; optimization objective->

The calculation mode of (a) is as follows:

for the mean value

Sum of variances->

Is a situation based on a proportional distribution, so that the mean +.>

Sum of variances->

Calculated from the following formula:

_η is the standard deviation

Weights of (2); for triangle blur number ++>

The treatment method is the same as->

4. The workflow scheduling method for reliability constraint in cloud environment as recited in claim 1, wherein for the estimated time t, the corresponding triangle ambiguity is

addition of the triangular fuzzy number:

comparison operation of triangle fuzzy numbers: if it is

Then->

The number multiplication operation of the triangle fuzzy number:

5. the workflow scheduling method for reliability constraint in cloud environment according to claim 1, wherein the genetic algorithm-based adaptive particle swarm algorithm is specifically as follows:

wherein ,

And fuzzy execution time->

And ∈h indicates the crossover operator and mutation operator, respectively:

Mutation operation is carried out, and the herb is added>

Will randomly change->

If r ₁ ≥p _m No variant behavior occurs;

wherein the two formulas are respectively corresponding toUpdating the human cognitive part and the social cognitive part, r ₂ and r₃ Is a random number of (0, 1), p _c For a given crossover probability, when r ₂ (or r) ₃ )<p _c When in use, then

wherein ,

indicating particle->

wherein ,

and

and

C is ₁ and c₂ Final value of (2).

6. The workflow scheduling method for reliability constraint in cloud environment of claim 5, wherein the specific cases of two particles to be compared are divided into the following three cases to be discussed:

F(P _i ^t )＝f(P _i ^t )

F(P _i ^t )＝Rel(P _i ^t )。

7. The workflow scheduling method for reliability constraint in cloud environment according to claim 1, wherein the step S3 specifically comprises: