FB-tree: a B+-tree for flash-based SSDs

Abstract

Due to their many advantages, flash-based SSDs (Solid-State Drives) have become a mainstream alternative to magnetic disks for database servers. Nevertheless, database systems, designed and optimized for magnetic disks, still do not fully exploit all the benefits of the new technology. We propose the FB-tree: a combination of an adapted B+-tree, a storage manager, and a buffer manager, all optimized for modern SSDs. Together the techniques enable writing to SSDs in relatively large blocks, thus achieving greater overall throughput. This is achieved by the out-of-place writing, whereby every time a modified index node is written, it is written to a new address, clustered with some other nodes that are written together. While this constantly frees index nodes, the FB-tree does not introduce any garbage-collection overhead, instead relying on naturally occurring free-space segments of sufficient size. As a consequence, the FB-tree outperforms a regular B+-tree in all scenarios tested. For instance, the throughput of a random workload of 75% updates increases by a factor of three using only two times the space of the B+-tree.