CN107995039B

CN107995039B - Resource self-learning and self-adaptive distribution method for cloud software service

Info

Publication number: CN107995039B
Application number: CN201711281436.2A
Authority: CN
Inventors: 陈星�; 林俊鑫; 项滔
Original assignee: Fuzhou University
Current assignee: Fuzhou University
Priority date: 2017-12-07
Filing date: 2017-12-07
Publication date: 2020-11-03
Anticipated expiration: 2037-12-07
Also published as: CN107995039A

Abstract

The invention discloses a resource self-learning and self-adaptive distribution method for cloud software service, which comprises the following steps: 1) establishing a QoS model of software service, and training the QoS model through a machine learning algorithm by using a data set of historical data as learning data; 2) constructing a fitness function of the software service resource allocation plan; 3) searching an optimal resource allocation plan based on a genetic algorithm, and performing online automatic allocation decision of resources. The invention has simple operation, can maintain the QoS value at a reasonable level, improves the resource utilization rate and has better cost performance.

Description

Resource self-learning and self-adaptive distribution method for cloud software service

Technical Field

The invention relates to the technical field of software engineering cloud computing, in particular to a resource self-learning and self-adaptive distribution method for cloud software service.

Background

Cloud computing is developing vigorously and a large number of cloud-based software services are apparent. To provide scalability and resiliency under varying workloads, cloud service providers are typically able to provide the ability to configure software and hardware resources in a shared infrastructure. The resiliency of the cloud has led us to manage the transition of cloud-based software service approaches. However, at the early stages of design, it is difficult for software engineers and cloud engineers to predict workload dynamics and runtime requirements for these cloud-based software services. The fact means that it becomes more and more complex for engineers to allocate appropriate resources for software services to guarantee good quality of service and low resource costs.

Software adaptation comes to bear against the complexity of current software systems. Such adaptive systems can configure and reconfigure themselves, enhance their functionality, continually optimize themselves, protect themselves and restore themselves, while keeping much of the complexity hidden from the user and administrator. Adaptive capabilities are needed in resource allocation for cloud-based software services because engineer intervention is difficult or even impossible. Traditional adaptive resource allocation methods are policy driven and are designed by experts. However, software services have different characteristics in workload type and resource preferences. Therefore, cloud engineers typically must develop a separate rule set for each system in order to efficiently allocate resources, resulting in high administrative costs and implementation complexity.

Machine learning, which allows computers to have the ability to learn themselves without explicit procedures, means that it allows systems to learn from data, and the autonomous learning ability needed to achieve enormous achievements in many areas is adaptive resource allocation, which is difficult to obtain because adaptive management is based on knowledge, experience, and rules. However, there are still two problems to be dealt with, firstly, the resource is configured with a large number of metrics and parameters. On the basis of machine learning, expert knowledge of resource allocation is difficult to define, and a proper model is trained from huge historical data; second, quality of service and low cost of resources need to be guaranteed when configuring resources for software services. Based on expert knowledge, it is difficult to make decisions on resource scheduling on-line.

Software services have different QoS values when the runtime environment changes. Depending on the nature of the originating agent, the environmental change may be classified as an external or internal change. External changes are initiated by external elements, while internal changes are applied by the management system.

TABLE 1

As shown in Table 1, there are three main elements in this problem area, including external changes, internal changes, and objects. External variations refer to workloads having different numbers and types. Internal changes refer to allocated resources that are made up of several virtual machines that vary in computing power and price. The object refers to an evaluation value calculated by a fitness function, which makes a trade-off between a QoS value and a resource cost. QoS may include goals for response time, data throughput, etc., whose values may be calculated by SLA contracts. The resource Cost (Cost) mainly comes from the lease Cost (CostL) and the stop Cost (CostD) of the virtual machine, and Cost is the Cost_L+Cost_D. Therefore, the goal of the adaptive system is to automatically allocate appropriate resources for software services, ensuring quality of service and resourcesLow cost of the source.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a resource self-learning and self-adaptive distribution method for cloud software service, which can maintain a QoS value at a reasonable level, improve the resource utilization rate and have better cost performance. In order to achieve the purpose, the technical scheme of the invention is as follows: a resource self-learning and self-adaptive distribution method for cloud software service comprises the following steps:

step 1: establishing a QoS model of the software service, and training the QoS model through a machine learning algorithm by using a data set of historical data as learning data:

Y＝Q(X) (1)

wherein the input matrix X comprises a number X of workloads_i,0Proportion of different types of task workload (x)_i,1,x_i,2,…,x_i,m) Allocated resource (x)_i,m+1,x_i,m+2,…,x_i,m+n) Wherein x is_i,1+x_i,2+…+x_i,m＝1，x_i,m+sThe number of the virtual machines of the s-th class is represented, wherein s is 1,2_i；

Step 2: constructing a fitness function of the software service resource allocation plan:

where fit is the fitness value, Cost_LIs the lease Cost of the virtual machine, Cost_DIs the stopping cost of the virtual machine, r₁、r₂Is weight, set according to the requirements of different systems;

and step 3: searching an optimal resource allocation plan based on a genetic algorithm, and performing online automatic allocation decision of resources. Further, the machine learning algorithm is a non-linear regression, support vector machine, or classification regression tree method. Further, the step 3 specifically includes:

step 31: binary coding is carried out on the resource allocation plan, each chromosome represents a resource allocation plan, and parameter values including the population size, the maximum iteration number and the chromosome number are initialized;

assuming that there are n types of virtual machines, the code length of the number of each type of virtual machine is v, the number of each type of virtual machine ranges from [0,2 ]^V-1]The ith chromosome in the tth iteration is defined as follows:

the s-th gene in the chromosome represents the number of the s-th class of virtual machines, and s is 1,2, …, n, and is described as follows by using a binary coding method:

step 32: calculating the fitness value of each chromosome according to the formula (2);

step 33: chromosomes with low fitness values are selected according to relative probabilities, and random mutation and cross operation are carried out on the chromosomes to update the population, wherein the relative probabilities are defined as follows:

step 34: calculating the fitness value of each chromosome of the new population;

step 35: repeating steps 33-34 until a finish loop condition is met;

step 36: and taking the chromosome with the lowest fitness value in the finally obtained new population as an optimal solution, and performing decoding operation on the chromosome to obtain the type and the corresponding number of the virtual machines, wherein the chromosome is the optimal resource allocation plan.

Further, the random mutation refers to randomly selecting a gene position in a chromosome and then mutating its attribute to a new value, i.e., 0 to 1, or 1 to 0.

Further, the crossover operation refers to the exchange of a gene for two parent chromosomes to generate a daughter chromosome.

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

(1) the method greatly reduces the search space when training the Qos prediction model;

(2) the invention can be used under most resource allocation conditions, because the QoS model and the fitness function can be adjusted according to different requirements;

(3) the invention can be independent of workload changes, i.e. predicting the workload and allocating resources independently, therefore, the invention can interact with the workload prediction model;

(4) the invention can maintain the stability of the Qos value at a reasonable level under the condition of the change of the operating environment, can improve the resource utilization rate and obtain higher cost performance.

Drawings

FIG. 1 is a self-learning and adaptive resource configuration framework;

FIG. 2 is a coding scheme for a chromosome of the present invention;

FIG. 3 is a schematic diagram of the operation of the gene mutation of the present invention;

FIG. 4 is a schematic diagram of the gene crossover operation of the present invention;

FIG. 5 is a QoS mapping formula in an embodiment of the present invention;

FIG. 6 is a schematic diagram of the variation of the average workload in an embodiment of the present invention;

fig. 7(a) is a schematic diagram of workload change when Swing is 200 in the embodiment of the present invention;

fig. 7(b) is a schematic diagram of the workload change when Swing is 500 in the embodiment of the present invention;

fig. 8(a) is a response time chart of a software service when Swing is 200 in an embodiment of the present invention;

fig. 8(b) is a Qos value map of the software service when Swing is 200 in the embodiment of the present invention;

fig. 9(a) is a response time chart of a software service when Swing is 500 in an embodiment of the present invention;

fig. 9(b) is a Qos value map of the software service when Swing is 500 in the embodiment of the present invention;

fig. 10(a) is the consumption and utilization of the CPU when Swing is 200 in an embodiment of the present invention;

fig. 10(b) shows memory consumption and usage when Swing is 200 in the embodiment of the present invention;

fig. 11(a) is the consumption and utilization of the CPU when Swing is 500 in an embodiment of the present invention;

fig. 11(b) shows memory consumption and usage when Swing is 500 in the embodiment of the present invention.

Detailed Description

The invention is further explained below with reference to the drawings and the embodiments.

Fig. 1 shows a framework for self-learning and adaptive resource configuration. The system mainly comprises a cloud resource module, an offline input module, an online self-learning module and a self-adaptive management module. The cloud resource module provides an api for monitoring and controlling cloud resources. The offline input module describes basic information about the workload, the virtual machines, and the fitness function. Based on the information of the offline input module, the online self-learning module trains the QoS model from historical data. The adaptive management module automatically allocates the appropriate software service resources using a genetic algorithm based on the QoS model.

The invention discloses a resource self-learning and self-adaptive distribution method for cloud software service, which comprises the following steps:

Y＝Q(X) (1)

wherein the input matrix X comprises a number X of workloads_i,0Proportion of different types of task workload (x)_i,1,x_i,2,…,x_i,m) Allocated resource (x)_i,m+1,x_i,m+2,…,x_i,m+n) Wherein x is_i,1+x_i,2+…+x_i,m＝1，x_i,m+sThe number of the virtual machines of the s-th class is represented, wherein s is 1,2_i(ii) a The data set of the historical data is shown in table 2:

TABLE 2

The machine learning algorithm is a non-linear regression, support vector machine, or classification regression tree method.

For the non-linear regression method, the regression equation was established as follows:

the mean squared error is used as a loss function, and the matrix W and the parameter b in the above equation are solved by least squares method parameter estimation.

For the Support Vector Machine (SVM) method, the hyperplane equation is set as follows:

parameter (u)^TV) mapping the feature space by a gaussian kernel, whose function is as follows:

for the classification regression tree (CART) method, the purity of the calculated dataset D ═ X, Y) is as follows:

wherein p is_kIs the proportion of the kth class in the dataset, wherein the dataset can be classified into r classes;

in the attribute column of the input matrix X, the Gini value calculation function of attribute att is as follows:

the one with the smallest kini coefficient is considered to be the optimal allocation attribute.

where fit is the fitness value, Cost_LIs the lease Cost of the virtual machine, Cost_DIs the stopping cost of the virtual machine, r₁、r₂Is weight, set according to the requirements of different systems; y is_iThe value of (c) can be calculated by the QoS model and the cost of the virtual machine can be calculated by the contract. Thus, for a given workload and assigned virtual machine, a fitness function value may be calculated.

And step 3: searching an optimal resource allocation plan based on a genetic algorithm, and performing online automatic allocation decision of resources. The step 3 specifically includes:

the value of each gene in the chromosome may be from 0 to 2^v-1 is unequal. For example, fig. 2 shows a chromosome encoded with a resource allocation plan of six VM types in a cloud environment, assuming that the maximum number of virtual machines of each type is 7 (i.e., v is 3), and thus the value of each gene is [0,7 ]]In the range of (1), the chromosome describes the scoreThe allocated resources include 3 vm1, 5 vm2, 5 vm3, 5 vm4, 6 vm5, and 7vm 7.

Step 32: calculating the fitness value of each chromosome according to formula (7);

step 35: repeating steps 33-34 until a finish loop condition is met;

The random mutation refers to randomly selecting a gene position in a chromosome and then mutating the attribute thereof to a new value, i.e., 0 to 1, or 1 to 0. As shown in FIG. 3, gene "101" was mutated to "010".

The crossover operation refers to the exchange of a gene for two parent chromosomes to produce daughter chromosomes. As shown in FIG. 4, "P1" and "P2" are two parent chromosomes. "s 1" and "s 2" represent two daughter chromosomes, in FIG. 4, the "010,101,101" of the gene is exchanged for the "011,100,110" in the gene "P2".

Experimental evaluation was performed by RUBiS benchmark test. RUBiS is an auction site prototype that mimics ebay.com and is commonly used to evaluate the performance scalability of application servers. It provides a client that can simulate user behavior in different workload patterns. The number of clients represents the number of workloads. Assuming that the number of workloads is typically in the range of [0,3000], there are two types of tasks in the workload. The experiment was conducted in a private cloud, with three types of virtual machines, as shown in table 3.

TABLE 3

Property	Small	Medium	Large
				CPU	1core	1core	1core
Memory device	1G	2G	4G
				Price(s)	1.761RMB/hour	1.885RMB/hour	2.084RMB/hour

In the present embodiment, the QoS value is mainly determined by the response time RT, which is mapped to a range of [0,1], as shown in fig. 5.

In this embodiment, the weight defining the fitness function is as follows:

collecting two-week operation data and respectively performing three machine learning method pairsThe QoS model is trained, and in order to evaluate the accuracy of the QoS model, a predicted value QoS is defined_predictAnd actual value QoS_actualThe tolerance E, confidence interval Pr and confidence level P are introduced as follows:

P＝P_r(QoS_actual-E≤QoS_predict≤QoS_actual+E) (12)

60 sets of tests were performed, the results are shown in Table 4,

TABLE 4

P	E＝0.1	E＝0.15
			NonlinearRegression	83.30％	86.70％
SVM	90.00％	96.70％
			CART	91.70％	95.00％

The confidence level of the QoS models based on the support vector machine and the decision tree both exceed 90% when the allowed error of the QoS is set to 0.1, and the confidence level of the QoS models based on the support vector machine and the decision tree both exceed 95% when the allowed error of the QoS is set to 0.15. The accuracy of the QoS model is acceptable from a system management perspective. For the QoS model, model training is carried out based on an original data set, and then a model with high accuracy is selected as a factor of a fitness function.

To evaluate the genetic algorithm based on-line decision method, the inventive method was compared to a rule based method, and table 5 describes a rule based scheme.

TABLE 5

Condition	Operation of
		Ucpu_ave<30％	Shutting down arbitrary virtual machines
Ucpu_ave>60％	Launching new VMs of any type

Two tests were performed simulating a three hour workload change as shown in figure 6, the workload average and the ratio of the two tasks each changed every half hour, and for each 30 minute period of the method of the invention the workload in the two tests would change in a sinusoidal manner with different values of swing as shown in figure 7(a) and figure 7 (b).

The inventive method and rule-based method are then used to allocate resources for software services in an automated manner, in accordance with workload changes. The resource allocation scheme when Swing is 200 is shown in table 6, and the resource allocation scheme when Swing is 500 is shown in table 7. Fig. 8(a) fig. 8(b) and fig. 9(a) the response time and QoS values of the software service are depicted in fig. 9 (b). Fig. 10(a) fig. 10(b) and fig. 11(a) fig. 11(b) describe the consumption and utilization of CPU and memory in detail.

TABLE 6

TABLE 7

As shown in fig. 8, when the workload varies in a sinusoidal pattern with a swing value of 200, the QoS values in the selected solution of the present invention are slightly lower than those in the rule-based solution, the average of the QoS values in the solution of the present invention is about 0.7, and the value of the rule-based solution is about 0.78. However, as shown in FIG. 9, when the workload varies in a sinusoidal pattern with a swing value of 500, the QoS values in the solution according to the invention are slightly higher than those in the solution based on rules, with an average value of about 0.75 for the QoS values and about 0.71 for the solution based on rules. Thus, there is not much difference between the inventive solution and the rule based QoS value, and the inventive solution is more stable than the rule based solution. It can be seen that the inventive solution enables the QoS value to be maintained at a reasonable level.

As shown in tables 6 and 7, the resource cost of the inventive solution is much lower than the rule-based resource cost under the condition of guaranteeing QoS premise. There are two main reasons: on the one hand, although the solution of the invention has almost less allocated resources than the rule-based ones, in practice the resources using both solutions are more or less the same, as shown in fig. 10 and 11. This means that the resource utilization of the inventive solution is almost higher than that of the rule-based solution. On the other hand, the three types of virtual machines had different cost performance, with the larger type being the best in this experiment, as shown in table 3. The method of the present invention tends to use large virtual machines, but the rule-based solution uses virtual machines randomly from among three types. It can be seen that the method of the present invention can provide a reasonable solution with a better performance/price ratio.

Although the present invention has been described with reference to the preferred embodiments, it is not intended to limit the present invention, and those skilled in the art can make variations and modifications of the present invention without departing from the spirit and scope of the present invention by using the methods and technical contents disclosed above. The above description is only a preferred embodiment of the present invention, and all equivalent changes and modifications made in accordance with the claims of the present invention should be covered by the present invention.

Claims

1. A resource self-learning and self-adaptive distribution method for cloud software service is characterized by comprising the following steps:

Y＝Q(X) (1)

and step 3: searching an optimal resource allocation plan based on a genetic algorithm, and performing online automatic allocation decision of resources; wherein, the step 3 specifically comprises:

step 35: repeating steps 33-34 until a finish loop condition is met;

2. The method of claim 1, wherein the machine learning algorithm is a non-linear regression, support vector machine, or classification regression tree method.

3. The method for resource self-learning and adaptive allocation according to claim 1, wherein the random mutation is to randomly select a gene location in the chromosome and then mutate its attribute to a new value, i.e. 0 to 1, or 1 to 0.

4. The method of claim 1, wherein the crossover operation is performed by exchanging a gene between two parent chromosomes to generate a child chromosome.