Tolerating Correlated Failures in Wide-Area Monitoring Services

Recently, there has been increasing interest in systems (e.g., Content Delivery Networks, Distributed Hash Tables, distributed file systems) running in large, wide-area distributed environments (e.g., Akamai, PlanetLab, RON). A key challenge to making these systems robust is the presence of correlated failures. In the context of a customizable wide-area monitoring service we developed, this paper is the first to comprehensively study the negative effects of failure correlation on availability and how to mitigate these effects. To achieve our availability goals, our service incorporates a replication design based on Signed Quorum Systems and a load-balancing design based on a novel database fragmentation algorithm POST. Through extensive live deployment on PlanetLab, trace-driven emulation and simulation, and a model-based sensitivity analysis, we observe that i) failure correlation is significant and has dramatic negative effects on availability; ii) correlation results in significantly diminishing returns for replication; and iii) our solutions are effective in mitigating the negative effects of correlation.