Eukv: Enabling Efficient Updates for Hybrid PM-DRAM Key-Value Store

Abstract

Key-value store, a fundamental component in a variety of modern data-intensive applications, can benefit from the emerging persistent memory (PM) to achieve high performance. Recent research proposes hybrid PM-DRAM key-value stores, which persist key-value pairs to PM in a log-structured approach for efficient storage and maintain a volatile hash index in DRAM for fast indexing. However, they fail to achieve high performance under update-intensive workloads due to the unawareness of PM’s characteristics. We propose Eukv, a hybrid PM-DRAM key-value store that supports efficient updates. Eukv (1) leverages out-of-place updates and lazy deletions to reduce small random PM reads and writes, (2) performs concurrency control only in a volatile hash index to further reduce small random PM reads and writes introduced by persistent lock operations, (3) uses cacheline-aligned writes to take full advantage of PM’s write performance, (4) improves garbage collection efficiency through hot and cold data separation, (5) employs multi-threaded garbage collection technology to reduce the impact of inefficient garbage collection on updates at high-capacity utilization. We conduct extensive evaluations on an 18-core machine equipped with Intel Optane DC Persistent Memory Modules. Experimental results demonstrate that Eukv obtains up to 5.6× speedup than the state-of-the-art hybrid PM-DRAM key-value store.