Cloud computing center green scheduling method considering double random fault influence

Abstract

The invention discloses a green scheduling method for a cloud computing center under the influence of network-power supply dual random faults based on an adaptive distribution estimation algorithm. By analyzing cloud computing scheduling characteristics under'network-power supply 'dual random faults, a three-objective optimization scheduling model for minimizing energy consumption and task completion time of a cloud computing center and maximizing a task success rate is constructed in combination with a fault management technology of task replication; secondly, designing four-dimensional probability matrixes such as a task processing sequence, a task processing position, a task processing speed and a backup task position so as to strengthen the optimization capacity of a scheduling algorithm; and a neighborhood search operator is designed for the three scheduling targets, so that the convergence speed of the algorithm is accelerated. Finally, algorithm parameters are determined through a Taguchi orthogonal experiment, a numerical simulation experiment is designed by using the parameters obtained through the orthogonal experiment, the effectiveness and superiority of the algorithm are analyzed by combining a C distance theory, and an optimization scheme is provided for a cloud computing center to give consideration to efficient operation, energy conservation and emission reduction.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud computing scheduling, and in particular to a green scheduling method for a cloud computing center under the influence of dual faults based on an adaptive distribution estimation algorithm. Background Art

[0002] As a computing environment, cloud computing has received widespread attention in recent years and has been applied to solve various problems in the consumer, technology and business fields. While cloud computing services are growing rapidly, the energy consumption of cloud computing data centers is also growing rapidly, and the high energy consumption of cloud computing needs to be solved urgently. In addition, with the continuous increase and expansion of hardware resources in cloud computing systems, the uncertainty of various resources is also increasing rapidly, resulting in frequent failures in the system during operation. Fault management of cloud computing centers is an important factor restricting the development of cloud computing technology. Existing technologies rarely pay attention to energy consumption and fault issues at the same time. Resource scheduling methods and key technologies that support green cloud computing in the context of fault management have important theoretical value and practical significance for reducing data center energy consumption and performing fault management.

[0003] The distribution estimation algorithm is a new type of optimization algorithm that has emerged in the field of evolutionary computing. It does not use crossover or mutation operations in genetics, but optimizes scheduling problems by establishing a probabilistic model that describes the problem and continuously updates and samples it. The application of the distribution estimation algorithm in actual production life shows that the algorithm can solve high-dimensional problems well and reduce time complexity. It has the advantages of strong global search capabilities and fast convergence speed, but its performance may be limited when dealing with complex problems. Summary of the invention

[0004] In view of the technical problems existing in the above background, the present invention provides a green scheduling method for cloud computing center under the influence of dual faults based on adaptive distribution estimation algorithm. Based on the green scheduling of dual faults in cloud computing center, an adaptive distribution estimation algorithm is designed to solve, and three adjustment operators are designed to realize neighborhood search and accelerate the convergence speed of the algorithm. It is helpful to manage the faults of cloud computing center and achieve the purpose of energy saving.

[0005] In order to solve the above problems, the present invention adopts the following technical solutions:

[0006] A green scheduling method for a cloud computing center under the influence of dual failures based on an adaptive distribution estimation algorithm, the method comprising the following steps:

[0007] Step 1: Green scheduling model in cloud computing environment based on dual failures of central network domain and power domain;

[0008] The tasks are scheduled using the following adaptive distribution estimation algorithm:

[0009] Step 2: Encode the scheduling model, initialize the probability matrix according to the priority constraint, perform probability matrix sampling, and generate a new population;

[0010] Step 3, use the improved operator to improve the diversity of TAS (temporary archive set), use TAS to update AS (archive set), and the highest η% of SLR (schedule length ratio) is used as the elite group update probability matrix;

[0011] Step 4, determine whether the end condition is met, if the condition is met, output the optimal green scheduling plan, if not, go to step 3 and continue the cycle until the end condition is met and the optimal plan is output;

[0012] Step 5: Taguchi orthogonal experiment is used to study the effect of parameters on algorithm performance at different parameter levels, find the best parameter combination, and compare the effectiveness and superiority of neighborhood search optimization design through C distance;

[0013] Furthermore, in step 1, the basic assumptions of the model are as follows: in the processing of the cloud computing center, the power area and the network area may fail independently. It is assumed that the failure probability of each power area follows a negative exponential distribution with parameter α1, and the maintenance time follows a negative exponential distribution with parameter δ1. It is assumed that the failure probability of each network area follows a negative exponential distribution with parameter α2, and the maintenance time follows a negative exponential distribution with parameter δ2; each task has only one backup task, and each task and its backup task can only be processed once; a processor can only process one task at a time; tasks on a processor are scheduled one by one in order; processor allocation is determined by the earliest finish time first (EFTF); tasks are not allowed to preempt;

[0014] Furthermore, in step 1, the green scheduling model established aims to minimize the maximum completion time and energy consumption and maximize the task completion rate; and ensure that each task must be processed once and can only be processed once; ensure that each task has only one backup; ensure that tasks on the processor can only be processed one by one; ensure that each task must start after its previous task is completed; the specific model and constraints are as follows:

[0015] min{C max ,TEC,-R}(1)

[0016]

[0017]

[0018] In the formula, C max is the maximum completion time, TEC is the energy consumption, R is the maximum task completion rate, so the goal is to minimize -R; n is the number of tasks to be processed, m is the number of processors; V jis the power supply voltage of task j, V low is the minimum power supply voltage, β is the energy consumption calculation constant, t j is the computation time of task j, rs j is the relevant speed of task j; x j,i,r = 1 means that task j is the rth task processed on machine i, otherwise x j,i,r =0;y j,i =1 means task j is backed up on machine i, otherwise y j,i =0;c j is the completion time of task j, M is a large constant, mj is the processor assigned to task j; δ(m i ,m j )=1 means mi≠mj, otherwise, m i =m j ; cm i,j is the communication time between task i and task j; i→j means task i is the immediate predecessor of task j.

[0019] Furthermore, for the calculation method of the maximum completion time, the start time and end time of task j on the processor are defined as follows:

[0020]

[0021] ET(j,m j )=ST(j,m j ,r j )+t j / rs j (5)

[0022] In the formula, ST(j,m j ,r j ) and ET(j,m j ) represents the start time and end time of task j on processor mj.

[0023] Furthermore, the calculation of the energy consumption of the cloud computing center includes the energy consumption during busy time and idle time. The energy consumption during busy time is significantly affected by the power supply voltage value and frequency of a task, and the energy consumption during idle time is greatly affected by the idle time. The specific calculation formula is as follows:

[0024]

[0025] E total =E busy +E idle (8)

[0026] In the formula, E busy is the energy consumption during peak hours, E idle is the energy consumption during idle time, Etotal is the total energy consumption; v i , t i and rs i are the power supply voltage, processing time and relative speed of task i, β is a constant parameter, and idle k To schedule the kth idle time, V low Indicates the minimum power supply voltage, d k Indicates the length of the kth idle space.

[0027] Furthermore, when calculating the task completion rate of the cloud computing center, a backup mechanism is used to avoid task failure in the event of a failure. When the region where the task is located fails and the region where its backup is located does not fail, the task can be restored through the backup mechanism. In this work, it is assumed that each task can have a backup. If task j is backed up on machine i, then y j,i =1, otherwise y j,i = 0. The effect of the fault resolution strategy is evaluated by simulation, and the formula is as follows:

[0028]

[0029] Where R is the task success rate under simulation results, S is the number of simulations, and scheme is the scheduling scheme of the problem instance; case i is a simulation instance under parameters α1, δ1, α2, δ2, i∈{1,2,3...,S}; In the simulation case i and scheduling scheme, processor k does not crash during the start to end time, otherwise u j =1 indicates that task j can be successfully executed, otherwise u j =0; p is the predecessor task of task j. The greater the task success rate, the more robust the scheduling plan.

[0030] Furthermore, in step 2, the scheduling model encoding includes task processing arrangement, task processor, task VSL allocation and backup location, which is expressed as follows:

[0031]

[0032] In the formula, i k Indicates that the task assigned to the kth position of the processor is i k , m k Representative Task i k Assigned to processor m k , Represents task i k Use relative speed To process, Representative Task ik Backup in processor superior.

[0033] Furthermore, in step 2, four probability models are used to iterate the offspring population:

[0034] (1) Probabilistic model of task arrangement

[0035] The task arrangement probability model generates task processing arrangements for the entire population. Due to the priority constraints between tasks, the relative position probability is designed as follows:

[0036]

[0037] where p i,j (g) represents the probability that task i is placed before task j in the g-th generation of permutation. Since task i is always before or after task j, p i,j (g)+p j,i (g) = 1. In addition, p 0,j (g) indicates that the virtual task is processed before task j. Initialize the probability model. Since the task processing arrangement should satisfy the priority constraint, p i,j The probability values ​​of are initialized as follows, and for unconstrained task pairs, their order is given randomly:

[0038]

[0039] (2) Probabilistic model of task location

[0040] The probabilistic model of task location is used to assign locations for tasks on processors. The design of LP is as follows:

[0041]

[0042] Among them lp j,l (g) represents the probability that task j is processed on processor l in the gth generation. The probability matrix is ​​initialized. At the beginning of the iteration, the probability of processing tasks on each computer is the same, so the probability matrix distribution is as follows:

[0043]

[0044] (3) Probabilistic Model of Dynamic Voltage Distribution

[0045] The probabilistic model of dynamic voltage allocation assigns voltage and relative speed to each machine:

[0046]

[0047] In the formula, vp i,v(g) indicates that the voltage-to-speed gear of machine i in the gth generation is v. Initialize the probability model and assume that the probability of each machine selecting the voltage-to-speed gear is the same at the beginning, so for each machine, the probability of selecting the four voltage-to-speed gears is 0.25.

[0048] (4) Probabilistic model of task backup location

[0049] The probabilistic model of task backup locations assigns each backup task a location on the processor:

[0050]

[0051] In the formula, bp j,l (g) indicates that task j is backed up on machine l in the gth generation. The probability model is initialized, and the probability of backing up tasks on each processor is equal at the beginning, and the probability matrix is ​​consistent with the task location probability matrix.

[0052] Furthermore, in step 3, the probability matrix is ​​updated as follows, and the completion time is normalized using the length of the critical path through the scheduling length ratio (SLR), as shown below:

[0053]

[0054] All individuals are sorted in ascending order relative to SLR, and then the top η% of solutions are selected as the elite population. The probability values ​​are updated by using the following population-based incremental learning method:

[0055]

[0056] Where τ is the learning rate, E is the size of the elite population, is the indicator function corresponding to the kth solution of the elite population. After the probability matrix is ​​updated, the updated probability model is sampled to generate new individuals. For the positions that have not yet been determined in the new individuals, an array of length n is given, and the lth element in the array represents the probability of task l being assigned to the position. If task l has been scheduled, the lth element is set to 0. Otherwise, by calculating p k,l The probability is obtained, k represents all tasks that have not been scheduled, and then the roulette method is used to normalize and sample the array.

[0057] Furthermore, in step 3, according to the idle conditions between tasks, three adjustment operators, namely the span adjustment operator, the energy consumption adjustment operator and the backup adjustment operator, are designed to optimize the three objective functions respectively.

[0058] Furthermore, in step 5, the C distance is used to compare the algorithm effects, and the dominance relationship between the two non-dominated sets A1 and A2 is reflected by comparing the quality of the solutions. The larger the C distance value, the more advantageous the algorithm is. The C distance calculation method is as follows:

[0059]

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The green scheduling method of the cloud computing center under the influence of dual faults based on the adaptive distribution estimation algorithm described in the present invention applies fault management technology and controls the energy consumption of the cloud computing center. Specifically, the replication technology is introduced as a fault management technology, the effect of the replication technology is characterized by constructing a task success rate function, and the maximization of the task success rate is added to the model as the objective function. In each example, the task success rate is calculated by simulation. A three-objective optimization model is constructed to describe the green scheduling problem of the cloud computing center under the influence of dual faults, with the goal of minimizing the energy consumption of the cloud computing center and the task completion time, and maximizing the task success rate, taking into account the constraint relationship between tasks and basic scheduling constraints. The scheduling strategy not only ensures the rapid completion of tasks, but also ensures the stability of task execution and the high efficiency of energy consumption.

[0062] An adaptive distribution estimation algorithm is designed to solve the scheduling problem. First, four probability models are designed to complete the description of the problem. In order to enhance the algorithm's optimization ability, neighborhood search is added to the algorithm. Three neighborhood search operators are designed to speed up the convergence of the algorithm, and a certain proportion of elite solutions are selected to update the probability matrix until the algorithm termination condition is reached. The effectiveness of the algorithm is verified through numerical experiments, and the effectiveness of the algorithm is compared from the perspective of C distance. Fast optimization and efficient execution of complex scheduling tasks are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to make the technical solution of the present invention clearer, the following briefly introduces the drawings required for describing the embodiments, wherein:

[0064] Figure 1 is the flow chart of the adaptive distribution estimation algorithm;

[0065] Figure 2 It is a DAG graph;

[0066] Figure 3 It is the span adjustment operator Gantt chart;

[0067] Figure 4 It is the Gantt chart of energy consumption adjustment operator;

[0068] Figure 5 is a schematic diagram of a backup adjustment operator;

[0069] Figure 6 is the orthogonal experimental response plot;

[0070] Figure 7 It is a box plot comparing EDA and EDA-NS;

[0071] Figure 8 It is the Pareto front diagram of EDA and EDA-NS;

[0072] Fig. 9 is the box plot comparing EDA-NS and NSGA-Ⅲ;

[0073] Fig.10 It is the Pareto front diagram of EDA-NS and NSGA-Ⅲ. DETAILED DESCRIPTION

[0074] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments, and the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that the embodiments described herein are only used to explain the present invention and are not used to limit the scope of protection of the present invention.

[0075] like Figure 1 As shown, it is a flowchart of the adaptive distribution estimation algorithm of this example, which includes the following steps:

[0076] Step 1: Green scheduling model in cloud computing environment based on dual failures of central network domain and power domain.

[0077] The basic assumptions of this step model are as follows: During the processing of the cloud computing center, the power area and the network area may fail independently. It is assumed that the failure probability of each power area follows a negative exponential distribution with parameter α1, and the maintenance time follows a negative exponential distribution with parameter δ1. It is assumed that the failure probability of each network area follows a negative exponential distribution with parameter α2, and the maintenance time follows a negative exponential distribution with parameter δ2; each task has only one backup task, and each task and its backup task can only be processed once; a processor can only process one task at a time; tasks on a processor are scheduled one by one in order; processor allocation is determined by the earliest finish time first (EFTF); tasks are not allowed to preempt.

[0078] The green scheduling model established aims to minimize the maximum completion time and energy consumption and maximize the task completion rate; ensure that each task must be processed once and can only be processed once; ensure that each task has only one backup; ensure that tasks on the processor can only be processed one by one; ensure that each task must start after its previous task is completed; the specific model and constraints are as follows:

[0079] min{C max ,TEC,-R}(1)

[0080]

[0081] In the formula, C max is the maximum completion time, TEC is the energy consumption, R is the maximum task completion rate, so the goal is to minimize -R; n is the number of tasks to be processed, m is the number of processors; V j is the power supply voltage of task j, V low is the minimum power supply voltage, β is the energy consumption calculation constant, t j is the computation time of task j, rs j is the relevant speed of task j; x j,i,r = 1 means that task j is the rth task processed on machine i, otherwise x j,i,r =0;y j,i =1 means task j is backed up on machine i, otherwise y j,i =0;c j is the completion time of task j, M is a large constant, mj is the processor assigned to task j; δ(m i ,m j )=1 means mi≠mj, otherwise, m i =m j ; cm i,j is the communication time between task i and task j; i→j means task i is the immediate predecessor of task j.

[0082] For the calculation method of the maximum completion time, the task flow running on the cloud computing center is abstracted into a DAG graph, which consists of n nodes and a set of edges, such as Figure 2 As shown in Figure 1, the entire application is executed from node 0. Each node represents a task, and the processing time of task i is t. i The directed edges define the dependencies between tasks. The edge (i, j) represents the priority constraint between tasks i and j. If there is a priority constraint between tasks i and j, task j cannot be processed before the data transfer is completed. The weight cm on the edge (i, j) is i,j Represents the communication time between two tasks. If the tasks with dependency files are assigned to the same processor, the data can be read from the memory and the communication time between tasks can be ignored. The start time and end time of task j on a processor are defined as follows:

[0083]

[0084] ET(j,m j )=ST(j,m j ,r j )+t j / rs j (5)

[0085] In the formula, ST(j,m j ,r j ) and ET(j,m j ) represents the start time and end time of task j on processor mj. The critical path (CP) is defined as the longest path in the directed acyclic graph without considering the communication time between tasks. The communication-computation ratio (CCR) is calculated as the average communication time divided by the average computation time. The larger the CCR, the more likely the application is communication-intensive. Otherwise, it is computation-intensive.

[0086] For the calculation of energy consumption of cloud computing center, DVFS technology is used to achieve energy saving goals. The cloud computing system consists of m fully connected processors in the data center. Each processor can only process one task at a time. Each processor supports DVFS and can run at different voltages. Four pairs of voltage and relative speed are provided in the present invention, which are randomly selected for each processor. In this work, a pair of voltage-relative-speed tables are set for all processors. The energy model of the system takes into account the energy consumption of the processor, including the energy consumption during busy time and idle time. The energy consumption during busy time is significantly affected by the power supply voltage value and frequency of a task, and the energy consumption during idle time is greatly affected by the idle time. The specific calculation formula is as follows:

[0087]

[0088] E total =E busy +E idle (8)

[0089] In the formula, E busy is the energy consumption during peak hours, E idle is the energy consumption during idle time, E total is the total energy consumption; v i , t i and rs i are the power supply voltage, processing time and relative speed of task i, β is a constant parameter, and idle k To schedule the kth idle time, V low Indicates the minimum power supply voltage, d k Indicates the length of the kth idle space.

[0090] When calculating the task completion rate of a cloud computing center, a backup mechanism is used to avoid task failure in the event of a failure. When the region where the task is located fails and the region where its backup is located does not fail, the task can be restored through the backup mechanism. In this work, it is assumed that each task can have a backup. If task j is backed up on machine i, then y j,i =1, otherwise yj,i = 0. The effect of the fault resolution strategy is evaluated by simulation, and the formula is as follows:

[0091]

[0092] Where R is the task success rate under simulation results, S is the number of simulations, and scheme is the scheduling scheme of the problem instance; case i is a simulation instance under parameters α1, δ1, α2, δ2, i∈{1,2,3...,S}; In the simulation case i and scheduling scheme, processor k does not crash during the start to end time, otherwise u j =1 indicates that task j can be successfully executed, otherwise u j =0; p is the predecessor task of task j. The greater the task success rate, the more robust the scheduling plan.

[0093] The tasks are scheduled using the following adaptive distribution estimation algorithm:

[0094] Step 2: Encode the scheduling model, initialize the probability matrix according to the priority constraint, perform probability matrix sampling, and generate a new population.

[0095] In this step, the scheduling model encoding includes task processing arrangement, task processor, task VSL allocation and backup location, which is expressed as follows:

[0096]

[0097] In the formula, i k Indicates that the task assigned to the kth position of the processor is i k , m k Representative Task i k Assigned to processor m k , Represents task i k Use relative speed To process, Representative Task i k Backup in processor superior.

[0098] Four probability models are used to iterate the offspring population:

[0099] (1) Probabilistic model of task arrangement

[0100] The task arrangement probability model generates task processing arrangements for the entire population. Due to the priority constraints between tasks, the relative position probability is designed as follows:

[0101]

[0102] where p i,j (g) represents the probability that task i is placed before task j in the g-th generation of permutation. Since task i is always before or after task j, p i,j (g)+p j,i (g) = 1. In addition, p 0,j (g) indicates that the virtual task is processed before task j. Initialize the probability model. Since the task processing arrangement should satisfy the priority constraint, p i,j The probability values ​​of are initialized as follows. For unconstrained task pairs, their order is given randomly.

[0103]

[0104] (2) Probabilistic model of task location

[0105] The probabilistic model of task location is used to assign locations for tasks on processors. The design of LP is as follows:

[0106]

[0107] Among them lp j,l (g) represents the probability that task j is processed on processor l in the gth generation. The probability matrix is ​​initialized. At the beginning of the iteration, the probability of processing tasks on each computer is the same, so the probability matrix distribution is as follows:

[0108]

[0109] (3) Probabilistic Model of Dynamic Voltage Distribution

[0110] The probabilistic model of dynamic voltage allocation assigns voltage and relative speed to each machine:

[0111]

[0112] In the formula, vp i,v (g) indicates that the voltage-to-speed gear of machine i in the gth generation is v. Initialize the probability model and assume that the probability of each machine selecting the voltage-to-speed gear is the same at the beginning, so for each machine, the probability of selecting the four voltage-to-speed gears is 0.25.

[0113] (4) Probabilistic model of task backup location

[0114] The probabilistic model of task backup locations assigns each backup task a location on the processor:

[0115]

[0116] In the formula, bp j,l (g) indicates that task j is backed up on machine l in the gth generation. The probability model is initialized, and the probability of backing up tasks on each processor is equal at the beginning, and the probability matrix is ​​consistent with the task location probability matrix.

[0117] Step 3, use the improved operator to improve the diversity of TAS (temporary archive set), use TAS to update AS (archive set), and the highest η% of SLR (schedule length ratio) is used as the elite group update probability matrix.

[0118] In this step, the probability matrix is ​​updated as follows, and the completion time is normalized using the length of the critical path through the scheduling length ratio (SLR), as shown below:

[0119]

[0120] All individuals are sorted in ascending order relative to SLR, and then the top η% of solutions are selected as the elite population. The probability values ​​are updated by using the following population-based incremental learning method:

[0121]

[0122] Where τ is the learning rate, E is the size of the elite population, is the indicator function corresponding to the kth solution of the elite population. After the probability matrix is ​​updated, the updated probability model is sampled to generate new individuals. For the positions that have not yet been determined in the new individuals, an array of length n is given, and the lth element in the array represents the probability of task l being assigned to the position. If task l has been scheduled, the lth element is set to 0. Otherwise, by calculating p k,l The probability is obtained, k represents all tasks that have not been scheduled, and then the roulette method is used to normalize and sample the array.

[0123] According to the idle situation between tasks, three adjustment operators, namely span adjustment operator, energy consumption adjustment operator and backup adjustment operator, are designed to optimize the three objective functions respectively.

[0124] like Figure 3As shown, a simplified span adjustment operator Gantt chart is given. First, find the idle time in the scheduling plan. For each idle time in the scheduling plan, consider the task after the idle time, that is, the shadow task J1 after the idle time on the M1 machine in the figure. Then find the parent task that delays the start time of J1, that is, J1 on M2. By adjusting the processing speed of the J1 task on M1, the processing time of the J1 task is shortened, so that the J1 task on M2 starts earlier and moves forward to the position shown in the dotted box in the figure below, thereby shortening the idle time and achieving the goal of minimizing makespan. Compare the energy consumption before and after accelerating J1. If the energy consumption does not increase, make the above adjustment.

[0125] like Figure 4 As shown in the figure, a simplified Gantt chart of the energy consumption adjustment operator is given. First, find the idle time in the scheduling plan. For each idle time in the scheduling plan, consider the task before the idle time, that is, the shadow task J2 before the idle time on the M2 machine in the example below. Without delaying subsequent tasks, try to slow down the speed of J2 to reduce energy consumption. After adjustment, the processing time of J2 is extended to the position shown in the red dotted box.

[0126] like Figure 5 As shown in , the backup adjustment operator mainly adjusts the location of the backup task. If a task and its backup task are in the same network domain or the same power domain, then the backup is likely to be invalid, which will reduce the task completion rate. For tasks whose processing location and backup location are in the same fault domain, this operator randomly assigns the backup location to a machine in a non-fault domain. Figure 5 ,Blue indicates that a task is placed on the processor, and orange indicates two situations of backup task placement.,When the orange task and the blue task are in the same power domain or the same network domain, the backup ,adjustment operator adjusts the backup task to the green shadow processor.,This adjustment can effectively avoid the failure of the original task and the backup ,task at the same time.

[0127] Step 4, determine whether the end condition is met, if the condition is met, output the optimal green scheduling plan, if not, go to step 3 and continue the cycle until the end condition is met and the optimal plan is output.

[0128] Step 5: Use Taguchi orthogonal experiment to study the impact of parameters on algorithm performance at different parameter levels, find the best parameter combination, and compare the effectiveness and superiority of neighborhood search optimization design through C distance.

[0129] In this step, the C distance is used to compare the algorithm effects. The dominance relationship between the two non-dominated sets A1 and A2 is reflected by comparing the quality of the solutions. The larger the C distance value, the more advantageous the algorithm is. The C distance calculation method is as follows:

[0130]

[0131] In this embodiment, the calculation of the objective function R needs to be simulated, and the number of simulations S is set to 500. The negative exponential distribution parameters α1 and α2 of the failure in each power area and network area are set to 20, and the negative exponential distribution parameters δ1 and δ2 of the failure maintenance time are also set to 20. The energy consumption coefficient is set to 1. The Taguchi orthogonal experiment is used to obtain the impact of parameters on the performance of the algorithm at different parameter levels. This embodiment selects the number of iterations, population size, learning rate, and the proportion of elite solutions used to update the probability model as four key parameters, and sets three levels for each parameter. The orthogonal experiment response diagram is shown in the figure. Figure 6 As shown in the figure, the population size is the most significant parameter. If the population size is too small, the solution space cannot be fully sampled, but when the population size is too large, the algorithm converges slowly. In addition, the learning rate is the second most significant parameter. A learning rate that is too large will lead to over-convergence, while a learning rate that is too small will affect the efficiency of the algorithm and lead to slow convergence. The impact of the number of iterations on the algorithm effect is that as the number of iterations increases, the quality of the solution obtained is higher. For the proportion of elite solutions, although it has little effect on the algorithm, an appropriate proportion helps the algorithm to reasonably update the probability model.

[0132] like Figure 7 As shown in the figure, the box plot of EDA and EDA-NS is compared with the results of the distribution estimation algorithm EDA-NS with neighborhood search and the ordinary EDA. According to the average value of C distance under different examples, after adding neighborhood search, the optimization ability of EDA is significantly strengthened, and the quality of the solution is higher.

[0133] like Figure 8 As shown in Figure 1, the Pareto front diagrams of EDA and EDA-NS, and the Pareto front diagrams of the two algorithms in the case of m=2 and n=5. It can be seen that although the number of non-dominated solutions obtained by the two algorithms is not much different, the quality of the solution obtained by EDA-NS is better.

[0134] like Fig. 9 As shown in the figure, the box plot of EDA-NS and NSGA-Ⅲ is compared with the results of EDA-NS and NSGA-Ⅲ. According to the comparison results of C values ​​under different examples, as the scale of the example gradually increases, EDA-NS shows more and more obvious advantages.

[0135] like Fig.10As shown in the figure, the Pareto front graphs of EDA-NS and NSGA-Ⅲ, the Pareto front graphs of the two algorithms in the case of m=4 and n=12. Although the solutions obtained by NSGA-Ⅲ show strong advantages in quantity and distribution, the quality of the solutions is poor. It often ignores other goals because of excessive pursuit of the minimum value of one goal, resulting in some inferior solutions. However, EDA-NS has stronger optimization ability, the solution quality is higher, and it is more valuable for reference. The two algorithms have their own advantages, but in the application of practical problems, EDA-NS is obviously a better choice.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A green scheduling method for cloud computing center under the influence of dual failures based on adaptive distribution estimation algorithm, characterized in that: The steps include: Step 1: Green scheduling model in cloud computing environment based on dual failures of central network domain and power domain; The tasks are scheduled using the following adaptive distribution estimation algorithm: Step 2: Encode the scheduling model, initialize the probability matrix according to the priority constraint, perform probability matrix sampling, and generate a new population; Step 3, use the improved operator to improve the diversity of TAS (temporary archive set), use TAS to update AS (archive set), and the highest η% of SLR (schedule length ratio) is used as the elite group update probability matrix; Step 4, determine whether the end condition is met, if the condition is met, output the optimal green scheduling plan, if not, go to step 3 and continue the cycle until the end condition is met and the optimal plan is output; Step 5: Use Taguchi orthogonal experiment to study the impact of parameters on algorithm performance at different parameter levels, find the best parameter combination, and compare the effectiveness and superiority of neighborhood search optimization design through C distance.

2. According to claim 1, a green scheduling method for cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm is characterized in that: In step 1, the basic assumptions of the model are as follows: During the processing of the cloud computing center, power areas and network areas may fail independently. It is assumed that the failure probability of each power area follows a negative exponential distribution with parameter α1, and the maintenance time follows a negative exponential distribution with parameter δ1. It is assumed that the failure probability of each network area follows a negative exponential distribution with parameter α2, and the maintenance time follows a negative exponential distribution with parameter δ2; each task has only one backup task, and each task and its backup task can only be processed once; a processor can only process one task at a time; tasks on a processor are scheduled one by one in order; processor allocation is determined by the earliest finish time first (EFTF); tasks are not allowed to preempt.

3. According to claim 1, a green scheduling method for cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm is characterized in that: In step 1, the green scheduling model established aims to minimize the maximum completion time and energy consumption and maximize the task completion rate; and ensure that each task must be processed once and can only be processed once; ensure that each task has only one backup; ensure that tasks on the processor can only be processed one by one; ensure that each task must start after its previous task is completed; the specific model and constraints are as follows: min{C max ,TEC,-R}(1) In the formula, C max is the maximum completion time, TEC is the energy consumption, R is the maximum task completion rate, so the goal is to minimize -R; n is the number of tasks to be processed, m is the number of processors; V j is the power supply voltage of task j, V low is the minimum power supply voltage, β is the energy consumption calculation constant, t j is the computation time of task j, rs j is the relevant speed of task j; x j,i,r = 1 means that task j is the rth task processed on machine i, otherwise x j,i,r =0;y j,i =1 means task j is backed up on machine i, otherwise y j,i =0;c j is the completion time of task j, M is a large constant, mj is the processor assigned to task j; δ(m i ,m j )=1 means mi≠mj, otherwise, m i =m j ; cm i,j is the communication time between task i and task j; i→j means task i is the immediate predecessor of task j.

4. According to claim 3, a green scheduling method for cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm is characterized in that: For the calculation method of the maximum completion time, the start time and end time of task j on the processor are defined as follows: ET(j,m j )=ST(j,m j ,r j )+t j / rs j (5) In the formula, ST(j,m j ,r j ) and ET(j,m j ) represents the start time and end time of task j on processor mj.

5. According to claim 3, a green scheduling method for cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm is characterized in that: The calculation of cloud computing center energy consumption includes energy consumption during busy time and idle time. The energy consumption during busy time is significantly affected by the power supply voltage value and frequency of a task, and the energy consumption during idle time is greatly affected by the idle time. The specific calculation formula is as follows: AND total =And busy +E idle (8) In the formula, E busy is the energy consumption during peak hours, E idle is the energy consumption during idle time, E total is the total energy consumption; v i , t i and rs i are the power supply voltage, processing time and relative speed of task i, β is a constant parameter, and idle k To schedule the kth idle time, V low Indicates the minimum power supply voltage, d k Indicates the length of the kth idle space.

6. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 3 is characterized in that: When calculating the task completion rate of a cloud computing center, a backup mechanism is used to avoid task failure in the event of a failure. In this work, it is assumed that each task can have a backup. If task j is backed up on machine i, then y j,i =1, otherwise y j,i = 0. The effect of the fault resolution strategy is evaluated by simulation, and the formula is as follows: Where R is the task success rate under simulation results, S is the number of simulations, and scheme is the scheduling scheme of the problem instance; case i is a simulation instance under parameters α1, δ1, α2, δ2, i∈{1,2,3...,S}; In the simulation case i and scheduling scheme, processor k does not crash during the start to end time, otherwise u j =1 indicates that task j can be successfully executed, otherwise u j =0; p is the predecessor task of task j. The greater the task success rate, the more robust the scheduling plan.

7. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 1 is characterized in that: In step 2, the scheduling model encoding includes task processing arrangement, task processor, task VSL allocation and backup location, which is expressed as follows: In the formula, i k Indicates that the task assigned to the kth position of the processor is i k , m k Representative Task i k Assigned to processor m k , Represents task i k Use relative speed To process, Representative Task i k Backup in processor superior.

8. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 1 is characterized in that: In step 2, four probability models are used to iterate the offspring population: (1) Probabilistic model of task arrangement The task arrangement probability model generates task processing arrangements for the entire population. Due to the priority constraints between tasks, the relative position probability is designed as follows: where p i,j (g) represents the probability that task i is placed before task j in the g-th generation of permutation. Since task i is always before or after task j, p i,j (g)+p j,i (g) = 1. In addition, p 0,j (g) indicates that the virtual task is processed before task j. Initialize the probability model. Since the task processing arrangement should satisfy the priority constraint, p i,j The probability values ​​of are initialized as follows, and for unconstrained task pairs, their order is given randomly: (2) Probabilistic model of task location The probabilistic model of task location is used to assign locations for tasks on processors. The design of LP is as follows: Among them lp j,l (g) represents the probability that task j is processed on processor l in the gth generation. The probability matrix is ​​initialized. At the beginning of the iteration, the probability of processing tasks on each computer is the same, so the probability matrix distribution is as follows: (3) Probabilistic Model of Dynamic Voltage Distribution The probabilistic model of dynamic voltage allocation assigns voltage and relative speed to each machine: In the formula, vp i,v (g) indicates that the voltage-to-speed gear of machine i in the gth generation is v. Initialize the probability model and assume that the probability of each machine selecting the voltage-to-speed gear is the same at the beginning, so for each machine, the probability of selecting the four voltage-to-speed gears is 0.

25. (4) Probabilistic model of task backup location The probabilistic model of task backup locations assigns each backup task a location on the processor: In the formula, bp j,l (g) indicates that task j is backed up on machine l in the gth generation. The probability model is initialized, and the probability of backing up tasks on each processor is equal at the beginning, and the probability matrix is ​​consistent with the task location probability matrix.

9. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 1 is characterized in that: In step 3, the probability matrix is ​​updated as follows, and the completion time is normalized using the length of the critical path through the scheduling length ratio (SLR), as shown below: All individuals are sorted in ascending order relative to SLR, and then the top η% of solutions are selected as the elite population. The probability values ​​are updated by using the following population-based incremental learning method: Where τ is the learning rate, E is the size of the elite population, is the indicator function corresponding to the kth solution of the elite population. After the probability matrix is ​​updated, the updated probability model is sampled to generate new individuals. For the positions that have not yet been determined in the new individuals, an array of length n is given, and the lth element in the array represents the probability of task l being assigned to the position. If task l has been scheduled, the lth element is set to 0. Otherwise, by calculating p k,l The probability is obtained, k represents all tasks that have not been scheduled, and then the roulette method is used to normalize and sample the array.

10. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 1 is characterized in that: In step 3, according to the idle situation between tasks, three adjustment operators, namely span adjustment operator, energy consumption adjustment operator and backup adjustment operator, are designed to optimize the three objective functions respectively.

11. The green scheduling method of cloud computing center under the influence of double failures based on adaptive distribution estimation algorithm according to claim 1 is characterized in that: In step 5, the C distance is used to compare the algorithm effects. The dominance relationship between the two non-dominated sets A1 and A2 is reflected by comparing the quality of the solutions. The larger the C distance value, the more advantageous the algorithm is. The C distance calculation method is as follows:

Images