Abstract
Due to the need of green computing and low cost, emerging paradigms such as edge and fog computing, and volunteer cloud have recently been introduced. In general, the volunteer cloud model targets globally distributed volunteer, highly heterogeneous, and non-dedicated machines. The inherent high degree of resource heterogeneity leads to varying levels of hardware and software failures and configuration faults on the unreliable and volatile volunteer hosts. As a result, the performance of deployed tasks is detrimentally impacted and is a key challenge, particularly in the case of scheduling algorithms. Most of the reputation models that have been used for reliability evaluation only evaluate the reliability of host machines by simply using the ratio of successfully completed tasks to total tasks requested. These models do not consider the resource utilization and the daily or weekly patterns of job behaviors or characteristics (e.g. priority of a job). Thus, the performance of tasks that run on these resources suffers and may take a substantial time to complete. Therefore, there is a need to proactively consider the reliability of host machines for effective and efficient management of resources in highly heterogeneous and distributed cloud environments. To address these challenges, this paper proposes a reputation and resource-based reliability model called ReMot. ReMot is an intelligent machine learning based model that utilizes historical data of the tasks and host machines to extract their resource usage patterns, in addition to other metrics such as task failure rate and resource utilization to predict the reliability of host machines. To validate ReMots approach, the researchers have utilized a large usage trace of real world applications made available by Google Inc. The results indicate that ReMot obtained more accurate reliability estimation than existing models and dynamically adapts to workload variations.