More for your money: exploiting performance heterogeneity in public clouds
Public cloud providers lease virtual machines(VM) on a server to customers. This model makes it easier for cloud providers to manage their infrastructure. However, on the other hand, it presents the customers with difficulty in terms of manually selecting the VM instance type. The inherent performance heterogeneity observed in public clouds further leads to the uncertainty of VM performance observed. We propose Jua, a customer facing system that solves the above problems. It consists of two main modules VM Instance Selector and VM placement Gamer. The former performs the task of choosing the appropriate VM instance type adopting a machine learning methodology and the latter module implements a number of placement gaming strategies to ensure that the customers get the best performance relative to the cost they incur. Initial evaluation of Jua confirms the viability of using a machine learning based approach for the selection of instance type. Finally, we develop a simulator to model and implement the various placement strategies and evaluate the same using a wide variety of synthetic benchmarks. We achieve a maximum performance improvement of 70%.