Pacman: An Efficient Compaction Approach for Log-Structured Key-Value Store...
summary
Pacman is a persistent-memory aware compaction design for log-structured key-value stores that offloads reference search to service threads, uses tagged pointers and DRAM-resident compaction metadata, batches flushes/fences and separates hot vs cold objects to cut PM write amplification. Implemented on state-of-the-art PM log-structured stores and benchmarked with YCSB and production traces, it raises compaction bandwidth by 2–4× and improves overall throughput by 1.5–1.8× under write-heavy workloads.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:33Persistent Memory (PM)
- 1:33Log-structured KV Systems
- 2:42Garbage Collection Overheads (1)
- 4:53High-latency Index Traversal
- 5:55Extra persistence work
- 6:33A large amount of data copying
- 6:57Excessive small random accesses
- 7:52Traverse Index with Shortcut
- 10:45Redesign Compaction Pipeline
- 12:20Separate Hot-Cold Data
- 14:13Reduce excessive PM accesses
- 15:58Experimental Setup
- 16:40Overall Performance
- 17:31Sensitivity Analysis
related talks
15mImproving Performance of Flash Based Key-Value Stores Using Storage Class Memory...
27mFPGA-Accelerated Compactions for LSM based Key-Value Store
35mDeveloping Software for Persistent Memory
29mSLM-DB: Single-Level Key-Value Store with Persistent Memory
18mPreview: KV-Stores and Databases
28mGearDB: A GC-free Key-Value Store on HM-SMR Drives with Gear Compaction
19mPolyStore: Exploiting Combined Capabilities of Heterogeneous Storage
22mPinK: High-speed In-storage Key-value Store with Bounded Tails