A trace-driven evaluation of cloud computing schedulers for IaaS
Zhihao Yao, Ioannis S. Papapanagiotou · 2017
Prior work in the Cloud domain has mainly focused on evaluating systems using synthetic workloads. Nonetheless, synthetic workloads are not always an accurate depiction of the real-world as they have been generated based on assumptions and statistical distributions that are not representative of the actual traffic that flows from the applications and microservices to the lower layers of the Cloud infrastructure. Moreover, the selection of diverse assumptions and distributions create additional complications on a fair comparison between platforms and reduces the ability to reproduce the experiment. In this paper, we compare synthetic workloads used in prior research with the Eucalyptus cloud traces. We find that the distributions employed in the synthetic workloads are not always in-line with the real usage pattern. Therefore, to deliver realistic and reproducible research outcomes, we propose a trace-driven research methodology and showcase an experimental design which employs a workload trace and simulation method. The experiment provides practical suggestions from a realistic input and environmental setting. In addition, it is straightforward to replicate the experiment so that the accepting process of research outcome is shorted by a practical implementation.