Alter: Towards Optimizing Persistent Indexes in Hybrid Memory Systems

Abstract

The rapid development of persistent memory (PM) provides a good opportunity to build a fast persistent index. Most existing PM-based index schemes store metadata on PM along with data. To ensure persistence and crash consistency, these schemes typically convert index updates into numerous small write operations to PM, resulting in low bandwidth utilization of PM devices.This paper presents a persistent index infrastructure Alter for a DRAM-PM hybrid memory system. By placing frequently accessed and updated metadata (keys in indexes) in DRAM and storing data (key-associated values) on PM, Alter avoids the small PM writes caused by metadata updates, thus taking full advantage of hybrid memory for high throughput. Alter consists of two modules. The log module records the write operations that are used to recover the index after a crash. Another is the metadata snapshot module, which records the instantaneous state of the index and can be used to accelerate recovery. We do experiments using both synthetic and real workloads. The results show that Alter-based indexes improve latency by up to 84% and multi-threaded throughput by up to 3.3x, compared with previous PM-based indexes. Moreover, Alter provides a common interface that makes it easy to implement different high-performance persistent indexes.