CN113867960B - Cloud load balancing hybrid model based on file types - Google Patents

Abstract

The invention provides a cloud load balancing hybrid model based on file types, which comprises the following steps of S1: classifying files in the cloud by using a support vector machine; s2: and inputting the classification result of the support vector machine into an ant colony optimization algorithm, and optimizing the load balancing performance of the cloud platform in the ant colony optimization algorithm in a multi-objective mode. The method is superior to the existing load balancing method, and has remarkable robustness and reliability on the cloud platform.

Description

Cloud load balancing hybrid model based on file types

Technical Field

The invention belongs to the field of cloud platforms, and particularly relates to a load balancing algorithm for a cloud platform.

Background

Today, cloud computing is playing an important role. Suppliers are offering premium services with SaaS, paaS, and IaaS, which have seen tremendous growth (about 21.5%) in the public cloud computing market over the past few years. Quality of service (QoS) also relates to other internal and external factors such as environmental problems, economics, sustainability, performance, energy consumption, new policies and development of new technologies. This means that the success of cloud computing depends largely on the effective support policies and intelligent decisions of suppliers and consumers. Similarly, other characteristics, such as load balancing, scalability, throughput, energy consumption, execution time, deadline constraints, optimizations, migration, and response time, are all considered by consumers and providers to maintain QoS. Since conventional algorithms cannot effectively solve QoS optimization problems, generic heuristic algorithms are used today to optimize QoS in the cloud.

Because of the huge and heterogeneous data volume in the cloud, more resources are needed to extract only relevant information. This situation becomes even more daunting when large-scale, computationally complex and resource-demanding applications need to be handled. In this case, data preprocessing may play an important role, where offline classification of data using a machine learning model may significantly reduce the execution time and memory requirements of the online processing stage. In addition, the tasks of the virtual machines also need to be reasonably allocated to ensure optimal load balancing.

Disclosure of Invention

The invention is made to solve the technical problems existing in the prior art, and aims to provide a novel efficient load balancing algorithm.

The invention utilizes file types to classify, provides a new hybrid multi-objective heuristic model on the basis of SVM and ACO, optimizes the model based on QoS indexes such as migration time, throughput time, overhead time, optimization time and the like, and ensures that the model can realize efficient load balancing in a cloud environment.

The invention provides a cloud load balancing hybrid model based on file types, which comprises the following steps,

s1: classifying files in the cloud by using a support vector machine;

s2: and inputting the classification result of the support vector machine into an ant colony optimization algorithm, and optimizing the load balancing performance of the cloud platform in the ant colony optimization algorithm in a multi-objective mode.

Further, the step S1 includes,

s11, introducing a kernel function to the support vector machine, and converting the original data space into a high-dimensional space containing a dot product transformation function, wherein the kernel function is as follows:

f () is an SVM function,representing a nonlinear function, u _i Representing support vector, alpha _i Represents Lagrange multiplier, u _j Representing member class labels, i, j representing the number of nodes, N representing the total number of nodes, and c representing the intercept;

s12 makes the data linearly separable using the following polynomial kernel function,

S(x,y)＝((x ^T y+1)) ^d

x is an input vector, y is a member class label, T is a conversion rank, d is a polynomial degree, and the polynomial degree is selected according to a learning algorithm.

Further, the step S2 includes,

s21, representing a virtual machine network in the form of an undirected weighted graph, wherein the virtual machine network is represented as an undirected graph G= (V, E), V represents a virtual machine or a node, E represents undirected edges with pheromone weights, and the pheromone weights represent overload and underload intensities between two nodes and are updated in the form of pheromones;

s22, initializing a pheromone, setting the initial pheromone to be 0.1, wherein the initial pheromone value is positioned between two nodes VMi and VMj, after the first iteration, the pheromone is globally updated, VMi represents an ith node, and VMj represents a jth node;

s23 calculating probability, ant k calculates probability of crossing edge by the following formulaTo decide to move from the current node VMi to the next node VMj,

n represents the number of neighbors of ant k, probability from node i to node jDepending on two parameters τ _ij And eta _ij ，τ _ij Representing pheromone, eta _ij Representing the possibility of movement from node i to node j, alpha and beta being used to control τ _ij And eta _ij Parameters of influence therebetween;

s24 updates the local pheromone using the following formula,

τ _ij pheromones representing nodes i through j, p is a constant pheromone evaporation coefficient as each ant passes through an edge ij,is the initial pheromone on edge ij;

s25, updating the global pheromone, calculating the global pheromone by adopting the following formula,

m represents the number of ants and is a number of the ants,is pheromone deposited on ij side of ant k in one iteration, L ^k Is the track t established by ant k _i Is a length of (2);

S26the values were calculated using the following formula:

t _i represented in machine m _j Task i running on;

s27, calculating the completion time, startTime by using the above formula _i Is the time at which task i is randomly allocated,is the estimated time to complete task i on machine j.

The method has the beneficial effect that a new load balancing hybrid model is provided on the cloud platform. The scheme uses a Support Vector Machine (SVM) to classify files in the cloud based on various file types (such as audio, video, text, images, etc.) in the cloud. The classification result of the SVM is further input into an ant colony optimization Algorithm (ACO), and a multi-objective optimization mode is adopted in the ACO to achieve better load balancing performance in the cloud. The model is superior to the existing method, and has remarkable robustness and reliability in a cloud platform.

Drawings

FIG. 1 illustrates a hybrid multi-objective heuristic model architecture proposed by the present invention;

FIG. 2 illustrates a virtual machine network;

FIG. 3 shows a performance comparison of a classifier;

FIG. 4 shows a performance comparison of the average number of SLAs violated;

FIG. 5 shows a performance comparison of average migration times;

FIG. 6 shows a performance comparison of average optimization time;

FIG. 7 shows a performance comparison of average throughput;

FIG. 8 shows a performance comparison of average overhead time;

Detailed Description

In order to obtain better classification results and effective task allocation, the invention provides a hybrid model based on a support vector machine and an ant colony algorithm, and compared with the existing model, the model has the best load balancing performance. By integrating these two models into one hybrid model with a multi-objective approach, their respective limitations are resolved and their overall benefits are enhanced. In addition, early studies focused on various factors such as cost, response time, and energy consumption by formulating appropriate single-target or multi-target QoS metrics. The development of general heuristic algorithms and hybrid heuristic algorithms is a new approach to solving such multi-objective optimization problems in cloud computing.

Hybrid heuristic algorithms are often used to solve classification, load balancing, fault tolerance, cost analysis, and energy conservation problems. However, classifying cloud data into various file types is a new contribution to the knowledge hierarchy. PostgreSQL and AWS have used classification methods to classify data. However, in cloud computing, classification of related data files (e.g., audio, video, text, images, maps) requires some additional effort to achieve accurate classification and to achieve load balancing. This problem can be solved in two steps. In a first step, a classification algorithm needs to be developed to perform accurate classification on the cloud dataset, resulting in accurate data classes. In a second step, the generated dataclasses are input into a load balancing algorithm, such as a generic heuristic algorithm. The SVM is a robust algorithm suitable for handling classification and regression problems. The ant colony optimization Algorithm (ACO) is one of the most widely used algorithms due to its good performance in handling load balancing problems. ACO has strong robustness, and can search the optimal solution faster. Because of its diversity, ACO has been widely used in various studies.

Therefore, in order to solve the above problem, the SVM is combined with the ACO, and a new hybrid multi-objective heuristic model is proposed on the basis of considering a plurality of important QoS indexes.

The invention provides a cloud load balancing hybrid model based on file types, which comprises the following steps,

s1: classifying files in the cloud using a Support Vector Machine (SVM);

s2: the classification result of the SVM is input into an ant colony optimization Algorithm (ACO), and a multi-objective optimization mode is adopted in the ACO to achieve better load balancing performance in the cloud platform;

further, the step S1 includes,

input: various types of files (video, audio, text, images, etc.);

and (3) outputting: a file category;

s11VM introduces a kernel function to transform the original data space into a high-dimensional space containing dot product transform functions. The kernel function is as follows:

f () is an SVM function,representing a nonlinear function, u _i Representing support vector, alpha _i Represents Lagrange multiplier, u _j Representing member class labels, i, j representing the number of nodes, N representing the total number of nodes, and c representing the intercept;

s12 since the input data is composed of various types of files, the invention selects POLY function in order to make the data linearly separable, i.e

POLY(u,v)＝((u ^k v+1)) ^s

s is the polynomial degree, and the polynomial kernel is defined as:

S(x,y)＝((x ^T y+1)) ^d

the polynomial degree must be selected according to a learning algorithm.

Further, the step S2 includes,

input: classified files;

and (3) outputting: data distribution;

s21 represents the virtual machine network in the form of an undirected weighted graph. The virtual machine network can be represented as an undirected graph g= (V, E), V representing virtual machines or nodes, E representing undirected edges with pheromone weights representing the intensity of overload and underload between two nodes and updated in the form of pheromones;

s22, initializing the pheromone, and setting the initial pheromone to 0.1. The initial pheromone value is located between two nodes, i.e. between VMi and VMj. After the first iteration, the pheromone is globally updated;

s23 calculating probability, ant k calculates probability of crossing edge by the following formulaTo decide to move from the current node VMi to the next node VMj,

representing the neighbors of ant k, probability from node i to node j->Depending on two parameters, namely pheromone tau _ij And the possibility of moving from node i to node j, with eta _ij Representing, alpha and beta are used to control tau _ij And eta _ij The influence of the two;

s24 updates the local pheromone. In the ant colony optimization algorithm, pheromones are updated locally and globally. The local pheromone is updated using the following formula,

τ _ij pheromone representing node i to node j, ρ being a constant pheromone evaporation coefficient, τ, when each ant passes through an edge ij _ij Is the initial pheromone on edge ij;

s25 updates the global pheromone. This occurs at the end of each iteration, when all ants construct a path. The global pheromone is calculated using the following formula,

m represents the number of ants and is a number of the ants,is the pheromone deposited on ij side by ant k in one iteration. L (L) ^k Is the track t established by ant k _i Is a length of (2);

s26 over time,the larger the value, the higher the pheromone on each side of the constructed path, and the following calculation formula is adopted:

s27 calculates the completion time using the above equation because the start time of the task depends on the task completion time previously assigned to the corresponding machine. StartTime when a machine is available _i Is the time at which task i is randomly allocated,is the estimated time to complete task i on machine j.

Further, the step S1 includes,

the invention combines a Support Vector Machine (SVM) with an ant colony optimization Algorithm (ACO) to provide a novel hybrid multi-objective heuristic model, and realizes high-efficiency load balancing in cloud computing. The model firstly classifies the input data and then carries out load balancing operation. This process begins by collecting data input in the form of video, text, audio, and images stored in a cloud environment. And classifying the data by using the SVM to obtain data class. And then load balancing is carried out on the data by using ACO.

The present invention proposes that the model first takes as input various files, such as audio, video, images and text from a cloud platform, and performs classification using a classifier. The algorithm iterates 100 times before assigning the data to the appropriate class. The high-dimensional complex data is processed by using a POLYSVM kernel. The output of the SVM classifier is a classified document.

And classifying the acquired data by using the SVM. The SVM introduces a kernel function that converts the original data space into a high-dimensional space containing a dot product transform function. The super function is as follows:

f () is an SVM function,representing a nonlinear function, u _i Representing support vector, alpha _i Represents Lagrange multiplier, u _j Representing member class labels, i, j representing the number of nodes, N representing the total number of nodes, and c representing the intercept;

since the input data is composed of various types of files, the invention selects POLY function, i.e. to make the data linearly separable

POLY(u,v)＝((u ^k v+1)) ^s .

s is the polynomial degree, and the polynomial kernel is defined as:

S(x,y)＝((x ^T y+1)) ^d

the polynomial degree must be selected according to a learning algorithm. When d=1, the linear kernel is satisfied. The polynomial kernel is applied to the curve in the input space.

Further, the step S2 includes,

suppose VM1, VM2,..vmn is a set of virtual machines, each of which is responsible for performing a task. Each task was performed 100 iterations and evaluated using the computational cost in time form. The ACO is used to compute a mapping of tasks on virtual machines, and the algorithm assigns each machine a task based on the available resources in the cloud environment.

The present invention represents a virtual machine network in the form of an undirected weighted graph. The virtual machine network may be represented as an undirected graph g= (V, E), V representing virtual machines or nodes, E representing undirected edges with pheromone weights representing the strength of overload and underload between two nodes and updated in the form of pheromones.

(1) Initializing pheromones

In the method proposed by the present invention, the initial pheromone is set to 0.1. The initial pheromone value is located between two nodes, i.e. between VMi and VMj. After the first iteration, the pheromone is globally updated.

(2) Calculating probability

Ant k calculates the probability of crossing the edge by the following formulaTo decide to move from the current node VMi to the next node VMj.

Representing the neighbors of ant k, probability from node i to node j->Depending on two parameters, namely pheromone tau _ij And the possibility of moving from node i to node j, with eta _ij Representing, alpha and beta are used to control tau _ij And eta _ij The effect of the above.

(3) Updating pheromones

The number of pheromones reflects the node type of the ant search. More pheromones on the path indicate that the target node is overloaded, so the ant will try to find another lesser path of pheromones, i.e. after encountering an overloaded node, it will find an underloaded node and assign tasks to that node.

In the ant colony optimization algorithm, pheromones are updated locally and globally. The local pheromone is updated using the following formula.

τ _ij Pheromone representing node i to node j, ρ being a constant pheromone evaporation coefficient, τ, when each ant passes through an edge ij _ij Is the initial pheromone on edge ij. The second level of pheromones is the global pheromone, which occurs at the end of each iteration, when all ants construct a path. The global pheromone is calculated using the following formula,

m represents the number of ants and is a number of the ants,is the pheromone deposited on ij side by ant k in one iteration. L (L) ^k Is the track t established by ant k _i Is a length of (c). Over time, the->The larger the value, the higher the pheromone on each side of the constructed path, and the following calculation formula is adopted:

the completion time is calculated using the above equation because the start time of a task depends on the task completion time previously assigned to the corresponding machine. StartTime when a machine is available _i Is the time at which task i is randomly allocated,is the estimated time to complete task i on machine j.

The meaning of the abbreviations in the present invention is described below.

QoS means quality of service.

ACO represents an ant colony optimization algorithm.

The SVM represents a support vector machine.

VM represents a virtual machine.

SLA indicates service level agreements

Example 1:

the present embodiment uses multiple types of files, including audio, video, image and text inventories, with equal proportions of each type of data. A total of 100000 data were used, 70% of which were used for training and the remaining 30% were used for testing. Each data type has a different file size, for example, an audio data set of 9Gb, a video data set of 19Gb, an image data set of 14Gb, and a text data set of 6Gb. The published data sets for UCI and YouTube were used throughout the experiment. Experiments were performed in CloudSim 4.0. A number of configurations are performed in CloudSim 4.0 for running simulations using resources such as data centers (2-16), hosts (2-32), virtual machines (5-1000), tasks (1000-14000), and task sizes (1 MB-1 GB). Table 2 shows the details of the data set used in this experiment.

Table 2 data samples

The classification accuracy test of this embodiment will be described below.

In order to check the accuracy of the proposed model of the present invention, the performance index given in table 3, which shows its average performance obtained from audio, video, image and text, was used. Performance indicators include accuracy, sensitivity, specificity, precision, recall, F-measure, and G-mean. It can be seen from table 3 that the proposed model of the present invention has the highest classification performance. This shows that the model classifies the files very accurately, which will have a huge impact on the scheduling.

TABLE 3 comprehensive results of Performance indicators

	Audio frequency	Video frequency	Picture picture	Text of	Average of
						Accuracy of	0.98	0.981	0.974	0.99	0.984
Sensitivity of	0.928	0.936	0.972	0.998	0.959
						Specificity (specificity)	0.99	0.991	0.945	0.998	0.982
Accuracy of	0.954	0.955	0.982	0.998	0.973
						Recall rate of recall	0.928	0.936	0.972	0.998	0.959
F-measure	0.94	0.945	0.977	0.998	0.966
						G-means	0.959	0.963	0.958	0.999	0.97
AUC	0.99	0.973	0.963	0.998	0.981

In this implementation, we selected Random Forest (RF), naive Bayes (NB), K-nearest neighbors (K-NN), and Convolutional Neural Network (CNN) for comparison with other well-known classifiers. Table 4 shows their classification properties. Compared with other classifiers, the method provided by the invention has better overall performance. The same effect can be seen in fig. 3, which shows the performance of this model compared to other models. The value is between 0,1, 0 representing zero precision and 1 representing the highest precision. Mathematically, the average accuracy of the model was 96.60%, followed by CNN, NB, 94.40%, RF, 93.60% and K-NN, 93.20%.

The load balancing test of this embodiment will be described below.

The model provided by the invention is evaluated by considering parameters such as execution time, migration times, optimization time, throughput time, overhead time and the like. In addition, the invention also compares the model with the following other methods:

ACOPS: a general heuristic hybrid algorithm combines the advantages of ACO and Particle Swarm Optimization (PSO) to solve the problem of virtual machine scheduling. The ACOPS algorithm predicts the workload of the virtual machine in cloud computing by adopting a dynamic scheduling strategy.

CPSO: diversity is improved and good global convergence is achieved while only cost is of concern. The disadvantage of this algorithm is that only one factor is considered, not multi-objective optimization.

QMPSO: a new hybrid heuristic algorithm that combines an improved PSO and an improved Q learning algorithm for load balancing in a cloud environment.

CSO: a general heuristic algorithm belongs to the group intelligent family and is based on the natural behavior of cats.

D-ACOELB: a heuristic algorithm based on an ACO algorithm is used for load balancing in cloud computing.

Figure 4 shows the average number of SLAs violated by the model in 14000 tasks drawn at random. Similarly, tasks are randomly selected to be 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 1000, 11000, 12000, 13000, and 14000. It was observed that the present invention proposes a model that rarely violates SLAs in different tasks. The stability of this model can be easily seen in fig. 4, while other models show a large number of violations in each task run. By taking into account SLA parameters such as performance, memory and CPU cycle usage, careful utilization of the virtual machine can effectively reduce violations. The same effect is to consume more energy and more time to optimize the solution, increasing computational complexity. Therefore, keeping the SLA as low as possible is one of the correct ways to achieve efficiency.

Figure 5 shows the model average migration time for 14000 tasks randomly extracted from 100000 data sets. It is observed that the present invention proposes that the migration time of the model to be completed on different tasks is minimal. The stability of this model can be easily seen in fig. 5, while on other models a longer migration time is required for each task run. Each time the memory size changes during operation affects the migration time of the virtual machine. As can be seen from fig. 6, other algorithms take up a lot of migration time and consume a lot of resources, because during execution, the resource requirements of the virtual machine may exceed the acquired resources, resulting in migration imbalance and non-scalability.

Fig. 6 shows the average optimization time for 14000 tasks randomly extracted from 100000 data sets. It was observed that the present invention proposes that the model optimizes itself as early as possible in different tasks. The stability of the model can be easily seen in fig. 6, which requires more time to optimize for each task run on other algorithms. Due to the advantage of high ACO convergence speed, the model can be helped to find the globally optimal solution in the earliest possible time. Fig. 8 demonstrates that the proposed model of the invention optimizes earlier than other algorithms. Furthermore, the inherent nature of the rapid optimization helps ACO solve complex problems with less computation time.

Fig. 7 shows the average throughput of 14000 tasks randomly extracted from 100000 data sets. It was observed that the proposed model shows maximum throughput over different tasks. This is because the earliest response time of the model helps to achieve faster throughput, while the response time of other algorithms is higher, resulting in low throughput.

Fig. 8 shows the average overhead time of 14000 tasks randomly extracted from 100000 data sets. It is observed that the present invention proposes a model with minimal overhead time on different tasks. The larger overhead may result in reduced performance and increased computational complexity of the system. Therefore, minimizing overhead is a better way to improve efficiency. ACO can significantly reduce overhead time due to its high probability and efficiency in finding globally optimal solutions.

The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (2)

1. The cloud load balancing hybrid model based on file types is characterized by comprising the following steps,

s1: classifying files in the cloud by using a support vector machine;

s2: inputting the classification result of the support vector machine into an ant colony optimization algorithm, and optimizing the load balancing performance of the cloud platform in the ant colony optimization algorithm in a multi-objective mode;

the step S2 includes:

s21, representing a virtual machine network in the form of an undirected weighted graph, wherein the virtual machine network is represented as an undirected graph G= (V, E), V represents a virtual machine or a node, E represents undirected edges with pheromone weights, and the pheromone weights represent overload and underload intensities between two nodes and are updated in the form of pheromones;

s22, initializing a pheromone, setting the initial pheromone to be 0.1, wherein the initial pheromone value is positioned between two nodes VMi and VMj, after the first iteration, the pheromone is globally updated, VMi represents an ith node, and VMj represents a jth node;

s23 calculating probability, ant k calculates probability of crossing edge by the following formulaTo decide to move from the current node VMi to the next node VMj,

n represents the number of neighbors of ant k, probability from node i to node jDepending on two parameters τ _ij And eta _ij ，τ _ij Representing pheromone, eta _ij Representing the possibility of movement from node i to node j, alpha and beta being used to control τ _ij And eta _ij Parameters of influence therebetween;

s24 updates the local pheromone using the following formula,

τ _ij pheromones representing nodes i through j, p is a constant pheromone evaporation coefficient as each ant passes through an edge ij,is the initial pheromone on edge ij;

s25, updating the global pheromone, calculating the global pheromone by adopting the following formula,

m represents the number of ants and is a number of the ants,is pheromone deposited on ij side of ant k in one iteration, L ^k Is the track t established by ant k _i Is a length of (2);

S26the values were calculated using the following formula:

t _i represented in machine m _j Task i running on;

s27, calculating the completion time, startTime by using the above formula _i Is the time at which task i is randomly allocated,is the estimated time to complete task i on machine j.

2. The file type based cloud load balancing hybrid model of claim 1, said step S1 comprising,

s11, introducing a kernel function to the support vector machine, and converting the original data space into a high-dimensional space containing a dot product transformation function, wherein the kernel function is as follows:

f () is an SVM function,representing a nonlinear function, u _i Representing support vector, alpha _i Represents Lagrange multiplier, u _j Representing member class labels, i, j representing the number of nodes, N representing the total number of nodes, and c representing the intercept;

s12 makes the data linearly separable using the following polynomial kernel function,

S(x,y)＝((x ^T _y +1)) ^d

x is an input vector, y is a member class label, T is a conversion rank, d is a polynomial degree, and the polynomial degree is selected according to a learning algorithm.