Internet-scale storage systems under churn - A steady state analysis

Content storage in a distributed collaborative environment uses redundancy for better resilience. This redundancy is either achieved by pure replication or using erasure codes for more efficient utilization of available storage. Increasingly such schemes are employed in peer-to-peer environments, though more realistically in local or wide area network settings, which is still characterized by moderate attrition (churn) of network membership. An mportant aspect of guaranteeing persistence and availability in such a system is to maintain the redundancy. A proactive mechanism of replenishing lost redundancy can however be prohibitive, and hence lazier maintenance schemes are desirable. Such systems' performance has typically been studied based on simulations, and more recently some prototype based experimental results have also been reported. While such a methodology is essential in designing and deploying a reliable system, the missing link in the existing literature is a sufficiently accurate analytical model to capture the dynamics of the system. The existing analytical models essentially capture the static resilience of the system. What's more realistic is to evaluate whether - given a rate of churn, and some chosen maintenance strategy, the system operates at a steady state and if so, what is the operational cost of such a system. Since simulations are restricted to chosen workloads (be it synthetic or trace driven), they are limited to only those chosen workloads, even while providing a good initial (hopefully empirical) insight to systems designers. A precise analytical model capturing the system's dynamicity however is expected to provide a more objective insight, and is the primary contribution of this work.