| Autor: |
Jin, Huixia, Fu, Yuanyuan, Yang, Gelan, Zhu, Xiaoning |
| Předmět: |
|
| Zdroj: |
Multimedia Tools & Applications; Feb2020, Vol. 79 Issue 7/8, p5335-5353, 19p |
| Abstrakt: |
Cloud-computing technologies and their application are becoming increasingly popular, which improves both enterprises' and individuals' working efficiency while at the same time greatly reducing users' cost. Besides, the scale of cloud platform and its application are rapidly expanding. Yet it's a challenging task to effectively utilize resource and guarantee quality of services to users. The quality of cloud task scheduling algorithm plays a key role in it. For one thing, traditional rule-based scheduling algorithms like FCFS and priority-based always focus on the algorithm itself instead of considering characteristics of Virtual Machines (VMs) and task, finally leading to poor operation effect. For another, one carefully selects a set of features based on sample data and employs machine-learning algorithms to train a scheduling policy. This method has the following deficiencies: quality of manually selected sample features directly affects that of the scheduling algorithm; many effective scheduling algorithms are based on a large number of labeled samples; however, it is very difficult to acquire these samples in reality; trained scheduling algorithms are often applicable only to specific environments and easy to be damaged. For the deficiencies of traditional scheduling algorithm and based on deep reinforcement learning (DRL) model, this paper presents a new-type model-free and end-to-end task scheduling agent which can interact with cloud environment and output the information of the virtual machine executing the task while inputting the original tasks of the cloud platform. The agent learns scheduling knowledge through the execution of tasks, and optimizes its scheduling policy. This algorithm completely solves the deficiencies of traditional scheduling algorithms like lower adaptability and flexibility, providing brand-new feasible solutions for task scheduling methods under cloud environments. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
Full text from SpringerLink