Performance-Predictable Resource Management of Container-based Genetic Algorithm Workloads in Cloud Infrastructure

Cloud computing, adopted by major providers like Amazon and Google, offers on-demand, pay-as-you-go services and resources through shared pools. Users submit workloads comprising multiple jobs, each containing tasks, including a specific genetic algorithm (GA) workload detailed in this thesis. This GA workload contains independent tasks from real-time multiprocessor allocation and Sudoku puzzle case studies, each with fixed deadlines and fitness requirements. Effective resource management is critical to enhance the Quality of Service (QoS) for cloud users. It involves resource allocation and adhering to QoS standards, guided by workload specifics. Container orchestration emerges as an essential deployment and management approach. This thesis focuses on managing multiple instances of genetic algorithms (GAs) in a cloud environment to achieve user-defined fitness levels within specified deadlines. It presents various approaches to allocate GAs to cloud nodes and control their execution iteratively. Initially, it introduces approaches such as fitness tracking (FT), fitness prediction (FP), fitness-prediction-based linear regression (FPLR), and fitness prediction based on weighted least squares (FPWLS) for managing the workload. To enhance resource efficiency, the thesis also addresses node interference, allowing multiple tasks to share resources while minimizing their impact on each other. It proposes a weighted-based node interference approach, considering fitness levels and response times during iterations to optimize task allocation. The performance of these approaches was experimentally evaluated by testing two GA applications and comparing them against state-of-the-art container-based orchestration approaches. Thus, different approaches were compared considering the number of successful tasks which can be defined by the number of tasks executed on time and achieved the fitness required. Comparison was also made between different approaches by taking iteration analysis into consideration. In situations where performance prediction was used, prediction errors like Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were used to evaluate and compare the performance of the prediction approaches.