Battle Royale deep reinforcement learning algorithm for effective load balancing in cloud computing

Battle Royale deep reinforcement learning algorithm for effective load balancing in cloud computing

연구 분야: Networking

학회: Cluster Computing

In recent times, cloud computing become a promising technology for providing computing sources and storage services to individual users and organizations over the internet via its service providers. However, one of the huge demanding situations in cloud computing is efficiently balancing the load throughout the available resources. The load balancing mechanism is required for allocating dynamic workloads equally among computing resources present in the cloud system to avoid virtual machines (VMs) being overloaded or underloaded. To overcome such issues, this paper proposes a hybrid Battle Royale Deep Reinforcement Learning (BRDRL) algorithm to enhance the efficiency of load balancing. The proposed algorithm is a combination of the two algorithms, namely Deep Reinforcement Learning (DRL) and Battle Royale Optimization (BRO). Cost, load balance, and makespan parameters are taken into account while doing effective load balancing, and the round-robin scheduling algorithm is used for scheduling tasks. The best-under loaded VM selection is determined using the BRO method, and job transfer is accomplished using DRL. The effectiveness of the proposed load balancing algorithm has been evaluated using the CloudSim tool. The proposed algorithm is analytically analyzed and outcomes are compared with the recent existing algorithms namely, Bald Eagle Assisted Butterfly Optimization Algorithm (BEABOA), Intercrossed Chimp and Bald Eagle Algorithm (ICBEA), Distributional Reinforcement Learning (DRL), Dynamic Q-Learning (DQL), Hybrid Dingo and Whale Optimization Algorithm (HDWOA), and Firefly and Honeybee-based Load Balancing Algorithm (FHLBA) in terms of throughput, response time, and makespan. The simulation outcomes reveal that proposed algorithm has performed 3.9%, 15.3% better in terms of throughput and response time respectively when compared to DRL. Further, it provides notably better outcomes than other hybrid optimization-based load balancing algorithms.