Navigating the Public Cloud Labyrinth: A Study of Price, Capacity, and Scaling Granularity Trade-offs

Public clouds offer numerous virtual machine (VM) types with complex trade-offs that tenants must navigate for cost-effective resource procurement. We identify three key qualitative axes using which these trade-offs can be succinctly expressed: price dynamism, offered capacity dynamism, and granularity of resource scaling. We offer several examples of contemporary VM offerings and where they fit within our scheme. Using our classification, we argue that related work lacks techniques for effectively navigating this trade-off space. We attempt to fill this gap by devising modeling and control techniques that combine Amazon EC2 style spot and on-demand instances with fine-grained resource allocation knobs that have begun to be offered by public clouds. We implement a Memcached prototype tenant that procures VMs from Amazon EC2 using our techniques. Using extensive evaluation based on real-world workload and spot price traces, we demonstrate the efficacy of our approach (e.g., 40% cost savings compared to a baseline while maintaining performance goals).