Improving Performance of Flash Based Key-Value Stores Using Storage Class Memory...
summary
Implementation uses Intel Optane DCPMM in AppDirect mode with memkind to extend memory and a RocksDB hybrid cache that applies a dram-first admission policy. Modified RocksDB places hot blocks in DRAM and colder blocks in SCM, evaluated on WhatsApp, Tectonic Metadata, and Laser traces. A 1P server with 64 GB DRAM plus SCM shows up to 80% throughput uplift, about 20% P95 latency reduction, 25–80% gains across configs, and lowers TCO to roughly 43–48% of a 2P platform.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:29Data Center Memory Requirements
- 1:24High-Performance Storage Servers at Facebook
- 2:34Storage Class Memory (SCM) Characteristics
- 3:01Intel Optane DC Persistent Memory (DCPMM)
- 3:59RocksDB Architecture
- 4:48Facebook RocksDB Workload Characteristics
- 5:42Intel Optane DC Persistent Memory Evaluations (WhatsApp)
- 6:56Facebook RocksDB extended with SCM
- 7:26Hybrid Cache Design and Components
- 8:16DRAM First Admission Policy
- 9:17Evaluation Setup and Implementations
- 9:58Workload Creation
- 10:11DRAM First Policy Throughput all workloads
- 10:44DRAM First Policy Latency improvement
- 11:39Deep Dive Analysis
- 12:13Cost and Performance
- 12:59Limitations of SCM Deployments
- 13:48Conclusion
- 14:20Acknowledgement
related talks
01h 07mEmerging Embdedded Memories and Storage
17mBreakthrough Data-Centric Computing with a New Memory Tier
25mFlashield: a Hybrid Key-value Cache that Controls Flash Write Amplification
24mPerformance tuning in BlueStore & RocksDB
31mReimagining the Future: Where the Market is Evolving
22mPinK: High-speed In-storage Key-value Store with Bounded Tails
35mDeveloping Software for Persistent Memory
06mEarly experience of Utilizing Persistent Memory for Database... (Work-in-Progress Report)