Introducing an improved deep reinforcement learning algorithm for task scheduling in cloud computing

In the cloud environment, task scheduling has always been a challenge. Failure to use a proper scheduling approach in cloud computing may cause high energy consumption and low resource efficiency. Due to the dynamism and limitation of cloud resources to execute diverse and time-varying requests of users, an effective scheduling mechanism is required to adapt to the dynamic conditions of the system. Deep learning, a practical method to deal with cloud computing resource management problems, has been noticed as an innovative idea in recent years. Among the deep learning algorithms used in this field, we can mention DQL, ERDQL and DRL-LSTM algorithms, which have achieved acceptable results in large data sets. But these algorithms have problems such as unrealistic reward estimation, long-term training, random sampling of memory, and not using the state space properly, which affect the results of their use in optimization problems. In this research, a deep reinforcement learning algorithm based on the development of PERDQL (prioritized experience replay DQL) is proposed for task scheduling in cloud computing to improve the problems of unrealistic reward estimation, long-term training, and random memory sampling. Also, the structure of the Markov decision process (MDP) is defined in the modeling part in such a way as to moderate the problems of not using the state space correctly. Therefore, by using the EPERDQN (extended prioritized experience replay deep Q-network) algorithm for scheduling tasks in cloud computing and defining the appropriate model based on the Markov decision process for it, the expected results lead to improvements in response time, waiting time, throughput, MakeSpan of tasks, and number of SLA violations. In this paper, the developed algorithm is evaluated with the GOCJ (Google Cloud Jobs) dataset, which reflects the real workload behavior of Google cluster traces and is suitable for research works in the cloud field. By conducting experiments with this dataset, the desired parameters are improved compared to other evaluated algorithms; for example, the throughput is 38% better than the DQN algorithm.