SILK: Preventing Latency Spikes in Log-Structured Merge Key-Value Stores
Talkat USENIX ATC '19 · on Thursday, 2019-07-11· at 17:15· USENIX ATC Track II: Grand Ballroom VII–IX·
summary
Implements an I/O scheduler inside RocksDB that allocates bandwidth opportunistically, prioritizes lower-level work and preempts higher-level compactions to avoid flash stalls. Evaluated on YCSB and a Nutanix production workload (500 GB dataset, 57% writes, 400 B values), the SILK prototype prevents level-0 full conditions, never stalls flushing, keeps throughput steady and reduces 99th‑percentile write latency by up to two orders of magnitude versus RocksDB and TRIAD.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 2:11What causes LSM latency spikes
- 4:40Write latency spikes
- 8:45What we learned
- 9:05SILK Scheduler
- 9:56SILK Principles
- 11:20Evaluation Results
- 14:10Conclusion
related talks
15mADOC: Automatically Harmonizing Dataflow Between Components in Log-Structured Key-Value
12mSpanDB: A Fast, Cost-Effective LSM-tree Based KV Store on Hybrid Storage
24mSplinterDB: Closing the Bandwidth Gap for NVMe Key-Value Stores
24mPerformance tuning in BlueStore & RocksDB
21mMatrixKV: Reducing Write Stalls and Write Amplification in LSM-tree Based KV Stores
11mEvolution of Development Priorities in Key-value Stores Serving Large-scale Applications
29mAccelerating Performance: Mastering RocksDB for High Speed Data Processing
51mSix New Important RocksDB Features And Planned Works