Abstract

The recent exciting development of persistent memory (PM) has led to many new proposals that directly operate and persist indexes on the memory bus, potentially removing the need for the storage stack. However, next-generation SSDs are quickly catching up with performance that overlaps with PM, effectively turning the storage hierarchy into a storage jungle. It is unclear how future persistent indexes (and data structures in general) should be designed, and more importantly, how their performance/cost would change given PM’s unconventional installation requirements compared to SSDs. This paper takes a first step to revisit the overall system cost and performance characteristics of the storage jungle, in the context of persistent indexes. We do so by experimentally evaluating PM and SSD indexes under real-world hardware constraints. We find that although PM has its own set of advantages, traditional DRAM-SSD hierarchies continue to be more cost-effective, and there is much to be further unleashed. Through careful analysis, we distill a series of observations, implications, and outlook on future index designs to navigate through the storage jungle.