Building workload-independent storage with VT-trees
Pradeep Shetty, Richard P. Spillane, Ravikant Malpani, Binesh Andrews, Justin Seyster, Erez Zadok · File and Storage Technologies · 2013
As the Internet and the amount of data grows, the variability of data sizes grows too--from small MP3 tags to large VM images. With applications using increasingly more complex queries and larger data-sets, data access patterns have become more complex and randomized. Current storage systems focus on optimizing for one band of workloads at the expense of other workloads due to limitations in existing storage system data structures. We designed a novel workload-independent data structure called the VT-tree which extends the LSM-tree to efficiently handle sequential and file-system workloads. We designed a system based solely on VT-trees which offers concurrent access to data via file system and database APIs, transactional guarantees, and consequently provides efficient and scalable access to both large and small data items regardless of the access pattern. Our evaluation shows that our user-level system has 2-6.6× better performance for random-write workloads and only a small average overhead for other workloads.