WO2016045515A1

WO2016045515A1 - Cloud task scheduling algorithm based on user satisfaction

Info

Publication number: WO2016045515A1
Application number: PCT/CN2015/089512
Authority: WO
Inventors: 蒋昌俊; 张亚英; 陈闳中; 闫春钢; 张冬冬; 陈熔仙
Original assignee: 同济大学
Priority date: 2014-09-23
Filing date: 2015-09-14
Publication date: 2016-03-31
Also published as: CN104333569A

Abstract

A cloud task scheduling algorithm based on user satisfaction, the algorithm mainly comprising several stages including normalising virtual machine and cloud task parameters; calculating Euclidean distance; selecting resources; updating Euclidean distance; calculating user satisfaction of a task; and calculating the resource usage costs of a single cloud task and the total costs of a system after execution of all tasks. The algorithm set forth in the present invention distributes tasks to the most suitable resources from the perspective of the user, thus better satisfying user requirements for various aspects including CPU, completion time, and bandwidth, and simultaneously effectively reducing the costs of resource usage by the user; in cloud computing, in which users are concerned with whether the costs paid reasonably match the service quality obtained, user requirements are highly satisfied. Compared to the prior art, the present invention provides an effective strategy for users to acquire the best service quality.

Description

Cloud task scheduling algorithm based on user satisfaction

Technical field

The invention relates to a cloud task scheduling algorithm.

Background technique

Cloud computing is a comprehensive development of parallel computing, distributed computing and grid computing. It is a commercial computing model that distributes tasks that previously required high-performance computers to be distributed on resource pools composed of a large number of inexpensive computers. The system is able to acquire computing power, storage space and information services as needed. However, while implementing these services, there is a problem to be considered. Different users have different requirements for the use of cloud computing resources, such as CPU, memory, completion time, bandwidth, usage fees, etc. How to adopt an effective strategy Let users get better service quality. The task scheduling algorithm of cloud computing is one of the ways to solve the above problems.

The traditional task scheduling algorithm pays attention to the efficiency of the server. For example, the task scheduling method aiming at the optimal completion time has better completion efficiency, but may result in high resource utilization rate and unbalanced system load. The equalization algorithm can provide an effective way to extend the bandwidth of network devices and servers, increase throughput, enhance network data processing capabilities, and improve network flexibility and availability. However, traditional task scheduling algorithms ignore the quality of service requirements of user tasks. , can not be a good allocation of resources on demand.

Summary of the invention

In the research process of cloud task scheduling technology, many classic scheduling algorithms are formed. They mostly consider the optimal completion time, minimum energy consumption, node load balancing, resource availability and reliability, and system from the perspective of cloud resource providers. Parameters such as utilization rate, and the algorithm proposed by the present invention focuses on the task completion time, the cost, the matching degree of cost and service quality, the satisfaction degree of the user's use of resources, and the load balancing of the system.

Cloud computing uses virtualization technology to encapsulate the underlying physical resources in the form of virtual machines, allowing virtual machines to perform user tasks. The scheduling problem is to map the user's tasks to resources with certain optimization goals. The cloud computing simplifies the matching of tasks and resources, so that the resources required by the tasks are reflected in the form of a virtual machine, so the search for resources is converted into Search for a virtual machine.

In order to implement the scheduling algorithm, the present invention first describes cloud tasks, virtual machines, and task classifications:

● The virtual machine is represented by a seven-tuple:

Vm _i =<id _i ,peNum _i ,ram _i ,bw _i ,C _cpu/num ,C _mem/MB ,C _bw/Mbps > (1)

The seven-tuple represents the virtual machine ID, CPU count, memory, bandwidth, and unit price of CPU, memory, and bandwidth.

● Cloud tasks are represented by eight-tuples:

t _i =<id _i ,type _i ,len _i ,exppe _i ,expram _i ,expbw _i ,s _i ,cost _i > (2)

The octet represents the cloud task ID, type, task size, number of CPUs expected, memory expected, bandwidth expected, user satisfaction of the task, and cost of executing the task.

Cloud task type: The present invention mainly considers the following QoS parameters:

a) Completion time: For the cloud task with real-time requirements, it needs to be completed in as little time as possible. What should be the two resources of CPU and execution speed.

b) Bandwidth: When cloud tasks require high communication bandwidth, such as multimedia streaming requirements, bandwidth requirements need to be prioritized.

c) Memory: When cloud tasks have high memory requirements, priority needs to be given to memory requirements.

For different cloud task requirements, user satisfaction is measured according to different QoS parameters. To this end, the present invention designs a weight vector, which represents the value recognition of the cloud platform for different resources, and uses the weight vector to adjust the selected virtual machine. The performance of the resource is better than the parameters to better improve the user's satisfaction with the use of resources. For example, for a real-time or time-sensitive cloud task, it is desirable to complete the task with a minimum completion time, so that a resource with strong computing power is required, so the CPU is given a larger weight. Let the weight vector of the i-th task be expressed as:

Ei=[ei ₁ ,ei ₂ ,ei ₃ ] (3)

Where ei ₁ , ei ₂ , ei ₃ correspond to the weights of the CPU, memory, and bandwidth, respectively, and

The idea of cloud task scheduling algorithm based on user satisfaction is as follows: For a given cloud task, select the cloud task with the highest priority in the system, normalize the parameters, and then normalize the cloud task. The cloud task and all the virtual machines in the system (all the virtual machines have been parameterized normalized) calculate the Euclidean distance. When calculating the Euclidean distance, according to the type of cloud task and the degree of recognition of the value of each parameter by the system, different degrees are given. The different weights of the performance parameters bind the current cloud task to the virtual machine with the lowest Euclidean distance. In order to balance the load of the system and ensure the completion time of all tasks, when the virtual machine is bound to a cloud task, update its Euclidean distance list, reducing the possibility that the next cloud task is assigned to the same virtual machine. After the current task is executed, calculate its user satisfaction and resource usage cost. After all tasks are executed, calculate the overall satisfaction of all cloud tasks and the total cost of the system.

The technical solution to be protected by the present invention is characterized as:

A cloud task scheduling algorithm based on user satisfaction, characterized in that it comprises the following steps:

Step 1. The resource parameters are normalized.

Normalize the performance parameters of tasks and virtual machines to the interval [0,1], so that the set X _ij ={X _1j ,...,X _tj }j is the number of performance parameters, which is the same performance parameter of the virtual machine. Collection, whose normalized value is:

GX _ij =(curX _ij -minX _ij )/(maxX _ij -minX _ij ) (5)

Where i is the number of tasks, j is the number of performance parameters, curX _ij is the current value of performance, minX _ij is the minimum of the same set of performance parameters, and maxX _ij is the maximum of the same set of performance parameters.

Step 2, calculate the Euclidean distance.

After the parameter normalization process, the parameter vector of the virtual machine is X={X ₁ , X ₂ , X ₃ }, and the parameter vector of the cloud task is Y={Y ₁ , Y ₂ , Y ₃ }. Considering three performance parameters of CPU, memory and bandwidth, the weight vector W={W ₁ , W ₂ , W ₃ } is obtained according to the type of cloud task. Then the formula for calculating the Euclidean distance is:

Where X _j represents the normalized parameter value of the jth resource in the i-th virtual machine; Y _j represents the expected value of the task for the j-th resource; W _j represents the weight of the j-th resource.

Step 3. Select a resource.

Each task selects the virtual machine with the smallest distance from its Euclidean task, and uses the party that controls the idle virtual machine. The load balancing is performed. Each virtual machine maintains a European distance table. When a task is successfully assigned to a virtual machine, the Euclidean distance table needs to be updated to increase the Euclidean distance between a certain type of resource and the virtual machine. , update formula is:

D _i ′=D _i (1+1/n), where n is the number of virtual machines (7)

Step 4, calculate user satisfaction.

After the task is completed, consider the completion of each task, including the completion time of the task and the user satisfaction of each task, and the overall satisfaction of all tasks. User satisfaction for a single task is:

Where s _i is the user satisfaction of task i; W _j is the weight of the performance parameter of the jth item; act _j is the actual consumption of the performance parameter of the jth item; exp _j is the user expectation of the performance of the j item of the cloud task value.

When 0≤|s _i |≤0.5, it is considered that the user is satisfied with the resource allocation of cloud task i; when 0.5<|s _i |≤1, the user is considered to be satisfied with the resource allocation of cloud task i; When s _i |>1, it is considered that the user is not satisfied with the resource allocation of the cloud task i; when the value of |s _i | is large, the user is considered to be very dissatisfied with the resource allocation of the cloud task i.

The overall user satisfaction for all cloud tasks is:

Where s _i is the user satisfaction of the i-th task; t is the number of cloud tasks. In a cloud computing system, the smaller the value of S, the higher the satisfaction of the services provided by all users of the system and the cloud computing service provider.

Step 5. Calculate the cost.

After performing each cloud task, calculate the cost of executing the task. The virtual machine charges the resources according to the unit. The total cost cost _i of the task consumption is:

Where P _i is the number of resources and C is the unit resource price.

After performing all cloud tasks, the total cost of the system is:

Where cost _i is the cost of the i-th task; t is the number of all cloud tasks. In a cloud computing system, the smaller the value of C, the lower the cost of executing all cloud tasks.

The algorithm of the invention allocates tasks to the most suitable resources from the perspective of the user, and better satisfies the requirements of the user on the CPU, the completion time, the bandwidth, and the like, and effectively reduces the cost of the user using the resources. In cloud computing, the concern for users is whether the cost paid and the quality of service obtained are reasonably matched, so that the user's needs are met to a higher degree. Compared with the prior art, the present invention provides an effective strategy for users to obtain better quality of service.

DRAWINGS

Figure 1 is a flow chart of a cloud task scheduling algorithm based on user satisfaction.

Figure 2 Simulation results of a cloud task scheduling algorithm based on user satisfaction.

Figure 3 shows the simulation results of the optimal completion time scheduling algorithm.

Figure 4 Task completion time comparison.

Figure 5 Comparison of task user satisfaction.

Figure 6 Task execution cost comparison.

detailed description

In the cloud computing, the user is not very concerned about the performance of the system. They are concerned with whether the cost paid and the quality of the service obtained are reasonably matched. The present invention proposes that the user's needs are satisfactorily satisfied. A cloud task scheduling algorithm based on user satisfaction. From the user's point of view, the algorithm assigns tasks to the most suitable resources, which better meets the user's requirements for CPU, completion time, bandwidth, etc., and effectively reduces the cost of users' resources. Finally, the present invention uses the CloudSim platform to simulate the algorithm and compares the algorithm with the currently used optimal completion time scheduling algorithm to verify the effectiveness of the algorithm in terms of user satisfaction and resource usage cost.

The present invention will be further described below in conjunction with an algorithm flow chart of the accompanying drawings.

Figure 1 is a flow chart of the algorithm of the present invention. As shown in the figure, the algorithm mainly includes parameter normalization of virtual machine and cloud tasks, calculation of Euclidean distance, resource selection, Euclidean distance update, user satisfaction of computing tasks, calculation of resource usage cost of a single cloud task, and execution of all tasks. After the completion of the system's total cost and other stages.

Step 1. The resource parameters are normalized. In order to more conveniently calculate the Euclidean distance, the present invention normalizes the performance parameters of the task and the virtual machine to the [0,1] interval, so that the set X _ij ={X _1j ,..., X _tj }j is a performance parameter. The number of the same type of performance parameters of the virtual machine, the normalized value is:

GX _ij =(curX _ij -minX _ij )/(maxX _ij -minX _ij ) (5)

Step 2, calculate the Euclidean distance. After the parameter normalization process, the parameter vector of the virtual machine is X={X ₁ , X ₂ , X ₃ }, and the parameter vector of the cloud task is Y={Y ₁ , Y ₂ , Y ₃ }. The present invention mainly considers three performance parameters of CPU, memory and bandwidth, and obtains a weight vector W={W ₁ , W ₂ , W ₃ } according to the type of the cloud task. Then the formula for calculating the Euclidean distance is:

The smaller the Euclidean distance between the task and the virtual machine, the better the user satisfaction can be obtained by selecting the virtual machine, and the cloud computing provider can better meet the value recognition requirements of different resources.

Step 3. Select a resource. Each task selects the virtual machine with the smallest distance from the Euclidean to perform the task, but considering the load balancing problem of the system, avoid all tasks being assigned to a very powerful virtual machine at the same time, and in order to optimize the task completion time as much as possible, Load balancing is performed by controlling idle virtual machines. Therefore, each virtual machine maintains a European distance table. When a task is successfully assigned to a virtual machine, it needs to update the Euclidean distance table to add certain resources and The Euclidean distance between the virtual machines, the update formula is:

D _i ′=D _i (1+1/n), where n is the number of virtual machines (7)

The resource selection process is as follows:

1.For

i=1 to m

2 Select VM by parameter of t _i to VM _i ; (all tasks with the same performance component vector)

3 For i=1 to t

4 For j=1 to 3

5 Compute GX _ij (parameter normalization);

6 For i=1 to t

7 Calculate the Euclidean distance D _i between the task and the virtual machine;

8 Select min D _i ;

9 Bind t _i to VM which has the min D _i ;

10 End;

Step 4, calculate user satisfaction. After the task is completed, you need to consider the completion of each task. This includes the completion time of the task and the user satisfaction of each task, as well as the overall satisfaction of all tasks. User satisfaction for a single task is:

The overall user satisfaction for all cloud tasks is:

Where s _i is the user satisfaction of the i-th task; t is the number of all cloud tasks. In a cloud computing system, the smaller the value of S, the higher the satisfaction of the services provided by all users of the system and the cloud computing service provider.

Step 5. Calculate the cost. After each cloud task is executed, the cost of executing the task needs to be calculated. The virtual machine charges the resources according to the unit. Therefore, the total cost cost _i of the task consumption is:

Where P _i is the number of resources and C is the unit resource price.

After performing all cloud tasks, the total cost of the system is:

Step 6, algorithm simulation. The simulated cloud environment of the present invention consists of five virtual machine nodes, a resource agent is established on the node, and 10 simulated user tasks are established. The simulation mainly considers the following aspects: the completion time of all cloud tasks, the user satisfaction of a single cloud task, the comprehensive satisfaction of all tasks, the execution cost of a single task, the cost of executing all cloud tasks in the system, and the resources in this simulation. The unit price is 1 (you can set it yourself when you use it). Figure 2 shows the simulation results of the cloud task scheduling algorithm based on user satisfaction. Figure 3 shows the simulation results of the optimal completion time scheduling algorithm.

As can be seen from Figure 4, in the optimal completion time algorithm, the final completion is task 3, the total completion time is 334.62ms, and in the cloud task scheduling algorithm based on user satisfaction, the final completion is task 7, the total completion The time is 389.92ms. The scheduling algorithm of the present invention is not as good as the classical optimal time scheduling algorithm when it is completed, but it is also relatively close.

It can be seen from Figure 5 that the satisfaction of most tasks and the total user satisfaction in the cloud task scheduling algorithm based on user satisfaction are better than the optimal completion time scheduling algorithm. This result reflects the effectiveness of the algorithm designed by the present invention.

It can be seen from Fig. 6 that the execution cost of most tasks in the cloud task scheduling algorithm based on user satisfaction is lower than the optimal completion time scheduling algorithm. The total system cost of the cloud task scheduling algorithm based on user satisfaction is 25587 units. The total system cost of the optimal completion time scheduling algorithm is 30,432 units. This result shows that the cloud task scheduling algorithm based on user satisfaction has an advantage in cost saving.

Through the above simulation results, compared with the optimal completion time task scheduling algorithm, the cloud task scheduling algorithm based on user satisfaction can better meet the needs of different users and improve users on the basis of ensuring good task completion time. Satisfaction, while also effectively saving implementation costs.

Innovation of the invention

1) Design a cloud task scheduling algorithm from the perspective of the user's satisfaction with the use of cloud computing resources.

2) The scheduling algorithm can better meet the users with different needs through the dynamic task allocation strategy under the condition of ensuring the completion time of the task, and can obtain good user satisfaction and overall system satisfaction, which can effectively save. The execution cost of the system.

Claims

A cloud task scheduling algorithm based on user satisfaction, characterized in that it comprises the following steps:

Step 1, normalization of resource parameters

Normalize the performance parameters of the task and the virtual machine to the [0,1] interval, so that the set X ij ={X ij ,...,X tj }j is the number of performance parameters, which is the same performance parameter of the virtual machine. Collection, whose normalized value is:

GX ij =(curX jj -minX ij )/(maxX ij -minX ij ) (5)

Where i is the number of tasks, j is the number of performance parameters, curX ij is the current value of performance, minX ij is the minimum of the same set of performance parameters, and maxX ij is the maximum of the same set of performance parameters.

Step 2, calculate the Euclidean distance

After the parameter normalization process, the parameter vector of the virtual machine is X={X 1 , X 2 , X 3 }, and the parameter vector of the cloud task is Y={Y 1 , Y 2 , Y 3 }; considering CPU and memory And three performance parameters of bandwidth, according to the type of cloud task to get the weight vector W = {W 1 , W 2 , W 3 }, then the formula for calculating the Euclidean distance is:

Where X j represents the normalized parameter value of the jth resource in the i-th virtual machine; Y j represents the expected value of the task for the j-th resource; W j represents the weight of the j-th resource;

Step 3, select a resource

Each task selects the virtual machine with the smallest distance from the Euclidean to perform the task, and uses the method of controlling the idle virtual machine to load balance. Each virtual machine maintains a European distance table, and when a task is successfully assigned to a virtual machine, After that, you need to update the Euclidean distance table to increase the Euclidean distance between a certain type of resource and the virtual machine. The update formula is:

D i ′=D i (1+1/n), where n is the number of virtual machines (7)

Step 4, calculate user satisfaction

After the task is completed, consider the completion of each task, including the completion time of the task and the user satisfaction of each task, and the overall satisfaction of all tasks; the user satisfaction of a single task is:

Where s i is the user satisfaction of task i; W j is the weight of the performance parameter of the jth item; act j is the actual consumption of the performance parameter of the jth item; exp j is the user expectation of the performance of the j item of the cloud task value;

When 0≤|s i |≤0.5, it is considered that the user is satisfied with the resource allocation of cloud task i; when 0.5<|s i |≤1, the user is considered to be satisfied with the resource allocation of cloud task i; When s i |>1, it is considered that the user is not satisfied with the resource allocation of the cloud task i; when the value of |s i | is large, the user is considered to be very dissatisfied with the resource allocation of the cloud task i;

The overall user satisfaction for all cloud tasks is:

Where s i is the user satisfaction of the i-th task; t is the number of all cloud tasks;

Step 5, calculate the cost

After each cloud task is executed, the cost of executing the task is calculated; the virtual machine charges the resource according to the unit, and the total cost cost i of the task consumption is:

Where P i is the number of resources and C is the unit resource price;

After performing all cloud tasks, the total cost of the system is:

Where cost i is the cost of the i-th task; t is the number of all cloud tasks.