Global-Scale Placement of Transactional Data Stores

Global-Scale Data Management (GSDM) empowers systems by providing higher levels of fault-tolerance, read availability, and efficiency in utilizing cloud resources. But, at which datacenters should data be placed? Current cloud providers offer tens of datacenters and hundreds of edge datacenters that are globally distributed all over the world. Unlike networks within a datacenter, the topology of theWide-Area Network (WAN) is asymmetric and diverse—the latency connecting a pair of datacenters can be an order of magnitude larger than the latency connecting another pair. This makes placement a significant factor in performance. However, it is not only placement. The specifics of the transaction management protocol play a crucial role in deciding which placement is ideal. In this paper, we develop GPlacer, a placement optimization framework that embeds the transaction protocol constraints into an optimization to derive both the data placement and the transaction protocol configuration that minimize the overall transaction latency. In developing GPlacer, we discover counter-intuitive lessons about data placement and transaction execution practices. Our evaluation shows that applying these lessons in addition to known best practices generate deployments that reduce the average transaction latency by up to 68%.