Design of Skiplist Based Key-Value Store on Non-Volatile Memory

Abstract

Non-volatile Memory is promising to persistent data storage, which has outstanding advantages against traditional storage devices such as HDD and SSD. One of its hugest advantages is its DRAM-like read latency and micro second level write latency, which is several hundred times faster than the original block device. However, one of the issues on using non-volatile memory as storage device is designing a suitable indexing system for data stores there, in which the characteristics of non-volatile memory can be able to make full use of. The state of the art indexing systems of non-volatile key-value stores are usually based on B+-tree or its variant, which are originally designed for mechanic hard disk and volatile memory. The semantics of B+-tree requires inside data being sorted and frequent split and merge for keeping its balance. However, both of sorting and splitting will cause extra write to non-volatile memory, which will downgrade the performance. As a result, B+-tree and its variant may not be naturally suitable for non-volatile memory. In this paper, we proposed a skiplist based indexing system for non-volatile memory key-value store(NV-Skiplist), which can take fully use of the features of both the non-volatile memory and DRAM. NV-Skiplist constructs its bottom layer in the non-volatile memory for data persistence and supporting range scan, it also builds its upper layers in the DRAM to retain fast index searching and prevent large consistent overhead. We also introduced a multi-ranged variant to increase the search performance. We evaluate the performance of NV-Skiplist on a non-volatile memory emulator with a server that has Intel Xeon E5-2620 v2 processor. The experimental results show that our design outperforms the original tree-based non-volatile key-value store on both insertion and search performance by both 27% and 12% in each case.