Extendable B-tree on Multi-channel Nonvolatile Memory Devices

B-tree is used to be a popular index structure in a variety of application scenarios such as file systems and databases. In a B-tree, each node is allocated to a fixed number of sectors of storage space, and once a node exhausts its allocated space, it overflows and has to be split. When this happens, the data items in the overflown node have to be migrated to the newly split node, which might incur severe performance overheads. The nodes of the B-tree are limited by the capacity upper bound to guarantee the worst-case time to search for a key in the node. However, nonvolatile memories (NVMs) are relatively slower on write operations, which makes the unsorted B-tree a preferred choice over the default B-tree. Observing the different performances and searching a key in a sorted node and unsorted node, we present how to adaptively extend the leaf nodes according to the key distributions to optimize the holistic latency to search a key in a B-tree. The main objective is to maximally postpone the timing of node overflows and improve the overall performance of the B-tree. The proposal in this study is verified through analytical and experimental studies, as the results are promising.