An artificial neural network based approach for energy efficient task scheduling in cloud data centers

Abstract

Energy efficiency is considered as a crucial objective in cloud data centers as it reduces cost and meets the standard set in green computing. Task scheduling an important problem becomes more complex and critical under energy efficiency consideration. Key issues in recent research on energy efficient task scheduling are execution overhead and scalability. Machine learning has been widely employed for energy efficient task scheduling problem but mostly used to predict resource consumption only instead of deciding the schedule itself. However, we used the neural network to decide which resource should be assigned to given task independently. In this paper, we proposed an energy efficient independent task scheduler using supervised neural networks with the aim to reduce makespan, energy consumption, execution overhead and number of active racks. Proposed artificial neural network-based scheduler takes incoming task and current cloud environment state as input and predict the best computing resource for given task as output which compiles our aim. We used genetic algorithm to generate a huge dataset (∼18 million training instances) and trained our neural network on this dataset using back propagation algorithm with 99.9% accuracy. We simulated experiments on heavily loaded and lightly loaded cloud environment and compared with well-known approaches: Genetic algorithm, MinMIN-MINMin heuristic and Linear regression based energy efficient task schedulers. Results clearly indicate that proposed work outperforms considered algorithms. In heavily (lightly) loaded environment, it improves makespan by 59% (64%), energy consumption by 45% (71%), execution overhead by 88% (43%) respectively and number of active racks by 70%.