Improving the performance of highly reliable storage systems

Highly reliable storage systems ensure protection of application data from loss. Reliability comes at the cost of expensive resources such as storage and bandwidth. In addition, challenges in data management are exacerbated due to the exponentially growing volumes of data in today's digital world. Content-addressable storage systems mitigate some of these problems by providing automatic mechanisms for detecting and deleting duplicate data in the storage infrastructure.;In this thesis, we provide several solutions that simplify the process of storage management. Specifically, we provide solutions that improve the scalability, performance and usability of highly-reliable content-addressable storage systems. We present an object partitioning technique called fingerdiff that improves duplicate elimination and scalability. Next, we present a block-level system called BLESS that enables primary storage applications such as file systems to continuously and transparently protect their data on a highly reliable content-addressable store. We evaluate BLESS using various workloads and demonstrate its efficacy in improving duplicate elimination, while reducing data protection costs.