4.1.

Fusion point design

By analysing the characteristics of genetic algorithm and ant colony algorithm, the fusion point design of these two algorithms is shown in Figure 1:

Figure 1.

Design of genetic ant colony fusion point.

In Figure 1, in the start time period T₀, T_a, genetic algorithm maintains a high evolution rate, but in T_b. After time, because genetic algorithm cannot effectively use heuristic information, its individual evolution rate begins to decrease gradually. On the contrary, the ant colony algorithm performs the following tasks in the time period T₀, T_a Due to the lack of early pheromones, the individual evolution rate is low, but in T_b After time, the individual evolution rate of ant colony algorithm is improved. Therefore, Ta is the best time point for the end and start of these two algorithms.

The algorithm steps are as follows:

4.2.

Transformation of initial pheromone

After the fusion point design is completed, the local optimal solution of the first algorithm needs to be converted into the initial pheromone of the second algorithm. The method adopted in this paper is: after the first algorithm runs the harness, the chromosome individuals with the highest fitness function value are screened out according to the set proportion (the set value here is 25%), the screened chromosomes are converted into the number of each resource point of cloud computing, and each resource point is converted into the pheromone path (m, n) of ants. The formula is as follows:

In equations (1) and (2), x_j represents the number of ants retained in resource point I by the individual with the jth genetic chromosome. θ represents pheromone conversion factor. The algorithm flow design in this paper is shown in Figure 2:

Figure 2.

Flow chart of fusion algorithm.

4.3.

Design of resource scheduling

The resource allocation of cloud computing is based on the reasonable allocation of N pending tasks uploaded by users to m computing resource nodes in the system under certain constraints. The design of resource allocation model requires that the task can be completed in a short time, and the power consumption of the system can also be kept at a low level. This paper designs a resource allocation model, which is specifically defined as follows:

Definition 1: task set X = {x₁, x₂, …, x_i, … x_n} is the task to be processed, n refers to the number of tasks, x_iis the ith task to be processed.

Definition 2: the calculation node set Y = {y₁, y₂, …, y_j, … y_т} is the calculation node included in the system, m refers to the number of calculation nodes, y_j refers to the jth allocable computing node.

Definition 3: T_exe(i, j) represents the expected time to complete task I on node j, using X_l(i) represents the data length of task I, Y_c(j) represents the execution rate of the computing node j. Then T_exe(i, j) can be expressed as:

Definition 4: T_d(i, j) represents the expected time required to transfer task I to computing node j. X_s(i) represents the amount of data that task I needs to transmit, Y_d(j) represents the data transmission rate of computing node j, then T_d(i, j) can be expressed as:

Definition 5: T_time(i, j) represents the expected time required to transfer the task I to the computing node j, and its value can be expressed as:

Since the resource allocation of cloud computing is parallel, each computing node completes its own work alone, then the expected time cost of the system to complete all tasks is T_c can be expressed as:

Power cost of completing resource allocation C_c can be expressed as:

In equation (7), C_e and C_d represents the power consumption of computing nodes in completing computing and transmission in unit time.

4.4

Improvement of ant colony algorithm

4.4.1

Improve Pheromone.

In order to reduce the ant colony algorithm falling into local optimization, we set the pheromone Volatilization Coefficient to achieve the adaptability of the ant colony algorithm. The adaptive pheromone coefficient volatilization formula is:

In formula (8), ρ(t) is the pheromone coefficient at time t. And ρ_min is the minimum value of the pheromone coefficient. The minimum value is set to avoid the algorithm convergence speed from falling too fast.

4.4.2

Design Load Balancing Adjustment Factor.

In order to avoid that too many tasks are assigned to the same node, resulting in overload of the node, a load balancing adjustment factor F is introduced here, as shown in equation (9):

The value of F in equation (9) can be calculated according to equations (3) and (4).

When the task X_i is assigned to the computing resource Y_j for execution, the pheromone of the computing resource Y_j is updated by the method of formula (10) here, and the other computing resources that are not assigned tasks are updated by the method of formula (11).

4.5

Design of genetic algorithm

Using genetic algorithm in resource allocation of cloud computing requires the design of fitness function in combination with the actual resource allocation. The larger the fitness function value, the more likely the genetic chromosome will be retained to the next generation. In order to achieve the premise of shorter task execution and completion time, we also need to ensure low power consumption of the system. The fitness function of completion time is designed Time and power cost fitness function Fitness_time and Fitness_power.

In order to find a balance between completion time and power cost, the fitness function of genetic algorithm is:

In equation (14), a represents the completion time weight value, and B represents the power consumption cost weight value. Between a and b, a + b=1, a & b ∈ 0, 1 is required.

5.1.

Experimental environment and parameter setting

The specific parameter settings of datacenter, virtual machine and cloudlet in cloud sim3.0 are shown in Table 1.

Table 1.

Experimental environment settings.

The parameter settings of fusion algorithm are shown in Table 2:

Table 2.

Parameter settings of this algorithm.

5.2

Experimental environment and parameter setting

This fusion algorithm is compared with the algorithms proposed in References^8-10. The comparison results of the three algorithms in terms of iteration times, time cost, power consumption cost, node load rate and so on are as follows:

Figure 3.

Comparison diagram of iteration times.

Figure 4.

Load balancing comparison chart.

Figure 5.

Comparison of time cost.

Figure 6.

Power consumption cost comparison.