Design Tradeoffs for Data Deduplication Performance in Backup Workloads

Data deduplication has become a standard component in modern backup systems. In order to understand the fundamental tradeoffs in each of its design choices (such as prefetching and sampling), we disassemble data deduplication into a large N-dimensional parameter space. Each point in the space is of various parameter settings, and performs a tradeoff among backup and restore performance, memory footprint, and storage cost. Existing and potential solutions can be considered as specific points in the space. Then, we propose a general-purpose framework to evaluate various deduplication solutions in the space. Given that no single solution is perfect in all metrics, our goal is to find some reasonable solutions that have sustained backup performance and perform a suitable tradeoff between deduplication ratio, memory footprints, and restore performance. Our findings from extensive experiments using real-world workloads provide a detailed guide to make efficient design decisions according to the desired tradeoff.