AB-Tree: A Write-Optimized Adaptive Index Structure on Solid State Disk

Abstract

Big Data boosts the development of data management and analysis in database systems but it also poses a challenge to traditional storages. Flash-based Solid State Disks (SSDs) are provided to deal with the new challenges brought by Big Data. However, SSD has the problem of read-write asymmetry due to the unique features of flash memory, which presents significant challenges in designing tree index for flash-based DBMS. In this paper, we designed the Adaptive Batched Tree (AB-Tree), a variant of the B-Tree to improve write performance. AB-Tree implements a bucket-based structure to perform bulk insertion. In addition, all the modifications to existing entries are performed in a lazy way so as to avoid small random writes. Besides, AB-Tree also has an adaptive bucket layout which can dynamically adapt to workload characteristics on-the-fly. Experimental results show that AB-Tree can achieve 6X to 133X gains over the state-of-art tree indexes across a range of workloads on SSDs.