Empowering bonobo optimizer for global optimization and cloud scheduling problem

Abstract

Task scheduling in cloud computing systems is an important and challenging NP-Hard problem that involves the decision to allocate resources to tasks in a way that optimizes a performance metric. The complexity of this problem rises due to the size and scale of cloud systems, the heterogeneity of cloud resources and tasks, and the dynamic nature of cloud resources. Metaheuristics are a class of algorithms that have been used effectively to solve NP-Hard cloud scheduling problems (CSP). Bonobo optimizer (BO) is a recent metaheuristic-based optimization algorithm, which mimics several interesting reproductive strategies and social behaviour of Bonobos and has shown competitive performance against several state-of-the-art metaheuristics for many optimization problems. Besides its good performance, it still suffers from inherent deficiencies such as imbalanced exploration-exploitation and trapping in local optima. This paper proposes a modified version of the BO algorithm called mBO to solve the cloud scheduling problem to minimize two important scheduling objectives; makespan and energy consumption. We have incorporated four modifications namely Dimension Learning-Based Hunting (DLH) search strategy, (2) Transition Factor (TF), (3) Control Randomization (DR), and 4) Control Randomization Direction in the traditional BO to improve the performance, which helps it to escape local optima and balance exploration-exploitation. The efficacy of mBO is initially tested on the popular standard CEC’20 benchmarks followed by its application on the CSP problem using real-world supercomputing workloads namely CEA-Curie and HPC2N. Results and observations reveal the supremacy of the proposed mBO algorithm over many contemporary metaheuristics by a competitive margin for both CEC’20 benchmarks and the CSP problem. Quantitatively for the CSP problem, mBO was able to reduce makespan and energy consumption by 8.20–23.73% and 2.57–11.87%, respectively against tested algorithms. For HPC2N workloads, mBO achieved a makespan reduction of 10.99–29.48% and an energy consumption reduction of 3.55–30.65% over the compared metaheuristics.