Closing the Performance Gap between DRAM and PM for In-Memory Index Structures

Emerging byte-addressable persistent memories such as Intel’s 3D XPoint enable new opportunities and challenges for in-memory indexing structures. While prior work has covered various aspects of persistent indexing structures, it has also been limited to performance studies using simulated memories ignoring the intricate details of persistent memory devices. Our work presents one of the first, in-depth performance studies on the interplay of real persistent memory hardware and indexing data structures. We conduct comprehensive evaluations of six variants of B+Trees leveraging diverse workloads and configurations. We obtain important findings via thorough investigation of the experimental results and detailed micro-architectural profiling. Based on our findings, we propose two novel techniques for improving the indexing data structure performance on persistent memories. Group flushing inserts timely flush operations improving the insertion performance of conventional B+-Trees by up to 24% while Persistency Optimized Log-structuring revisits log-structuring for persistent memories improving the performance of B+-Trees by up to 41%.