Dynamically resilient and agile fine-grained replication configuration

Service resiliency is crucial to the operation of global-scale cloud applications with fine-grained replication (FGR). If poorly designed, a few disruptions may cause cascading failures or severely impair the level of service. Due to high complexity of FGR configuration problem, existing techniques usually rely on over-provisioning and static replica configuration to achieve reliability to a certain extent. Such approaches may suffer from slow disruption response and unnecessary sacrifice on service availability. This paper proposes a novel problem formulation of online FGR configuration for service resiliency, which includes a new definition of service availability and a new constraint of zero-data-move configuration. To efficiently solve the proposed problem, this paper presents a brand new optimal dimension reduction technique - Replica Vector Decomposition (RVD), which greatly reduces problem complexity while preserving solution optimality (by formal proof). A RVD-based framework is designed for the online FGR configuration problem. Our evaluation results show that RVD-based FGR configuration scheme achieves substantial speedup of online disruption response, significantly improves service resiliency, and maintains a high level of service in the face of multiple disruptions.