DRLBTSA: Deep reinforcement learning based task-scheduling algorithm in cloud computing

doi:10.1007/s11042-023-16008-2

Return

DRLBTSA: Deep reinforcement learning based task-scheduling algorithm in cloud computing

2023-06-17

PRE

Sudheer Mangalampalli *

Ganesh Reddy Karri

Mohit Kumar

Osama Ibrahim Khalaf

Carlos Andres Tavera Romero

GhaidaMuttashar Abdul Sahib

DOI：10.1007/s11042-023-16008-2

Original

Original request for help

Save

Abstract

En 中文

Task scheduling in cloud paradigm brought attention of all researchers as it is a challenging issue due to uncertainty, heterogeneity, and dynamic nature as they are varied in size, processing capacity and number of tasks to be scheduled. Therefore, ineffective scheduling technique may lead to increase of energy consumption SLA violations and makespan. Many of authors proposed heuristic approaches to solve task scheduling problem in cloud paradigm but it is fall behind to achieve goal effectively and need improvement especially while scheduling multimedia tasks as they consists of more heterogeneity, processing capacity. Therefore, to handle this dynamic nature of tasks in cloud paradigm, a scheduling mechanism, which automatically takes the decision based on the upcoming tasks onto cloud console and already running tasks in the underlying virtual resources. In this paper, we have used a Deep Q-learning network model to addressed the mentioned scheduling problem that search the optimal resource for the tasks. The entire extensive simulationsare performed usingCloudsim toolkit. It was carried out in two phases. Initially random generated workload is used for simulation. After that, HPC2N and NASA workload are used to measure performance of proposed algorithm. DRLBTSA is compared over baseline algorithms such as FCFS, RR, Earliest Deadline first approaches. From simulation results it is evident that our proposed scheduler DRLBTSA minimizes makespan over RR,FCFS, EDF, RATS-HM, MOABCQ by 29.76%, 41.03%, 27.4%, 33.97%, 33.57% respectively. SLA violation percentage for DRLBTSA minimized overRR,FCFS, EDF, RATS-HM, MOABCQ by48.12%, 41.57%, 37.57%, 36.36%, 30.59% respectively and energy consumption for DRLBTSA over RR,FCFS, EDF, RATS-HM, MOABCQ by36.58%,43.2%, 38.22%, 38.52%, 33.82%existing approaches.

Keywords:

Cloud Computing

Task Scheduling

Machine Learning

Deep Q- Learning

Makespan

Energy consumption

SLA violation