Executing Locally to Reduce Cost and Latency for Consistent Applications

Abstract

High application latency has a significant effect on user traffic [23]. However, reducing latency while guaranteeing consistency is difficult because storage systems are located far from end users. While recent work improves latency for strong consistent applications by running closer to users, it imposes high cost for storage and compute [19]. In this project, we propose Radical-Local in order to minimize cost overhead while providing low latency for consistent applications. Radical-Local speculatively executes applications on end-user machines. It guarantees linearizability by simultaneously syncing with the primary datacenter. When speculative execution succeeds, Radical-Local avoids the cost of using cloud compute. It is also faster than execution in a distant datacenter. For real application deployments, Radical-Local reduces cost by 9 − 29% and median latency by 7% compared to previous literature [19]. This represents a latency improvement of 48% compared to execution in the datacenter.