Towards a leaner geo-distributed cloud infrastructure

Abstract

Modern cloud infrastructures are geo-distributed. Geodistribution offers many advantages but can increase the total cloud capacity required. To achieve low latency, geo-distribution forfeits statistical multiplexing of demand that a single data center could benefit from. Geo-distribution also complicates software design due to storage consistency issues. On the other hand, geodistribution can lower costs through eliminating redundancies at individual sites or exploiting regional differences in energy prices. We discuss several factors that influence geo-distributed capacity provisioning, and quantify latency, availability, and capacity trade-offs that emerge. We describe open research challenges in designing software that efficiently uses cloud capacity.