Panel: Flexible Data Placement (FDP)
An OCP Global Summit storage-track panel on NVMe Flexible Data Placement (FDP), the host-directed data-placement feature that reduces SSD write amplification (WAF) toward 1.0. Four vendor speakers — Javier Gonzalez (Samsung), Rory Bolt (Kioxia), Jonathan Hughes (Solidigm), and Jiajia (ScaleFlux) — cover FDP's maturity and Linux ecosystem support, real RocksDB and Linux-build WAF results, Solidigm's multiple reclaim group extension, and ScaleFlux's resource grant and isolation proposals.
Key Takeaways
- FDP is mature and deployed: ratified in NVMe Express, adopted into the OCP Datacenter NVMe spec since revisions 2.5/2.6, shipping from Samsung and others for 2-3 years, with Linux support across NVMe, block layer, SPDK, and the XFS/EXT4/Btrfs file systems.
- FDP's edge over prior attempts (open channel, streams, ZNS) is backwards compatibility: 70-80% WAF improvement is achievable through configuration alone without application changes.
- Kioxia's open-source U-Frog RocksDB plugin uses io_uring to bypass the file system and push drive WAF to ~1.0, lifting query throughput from ~8,000 to ~15,000 QPS; a no-code temp-FS/swap experiment cut Linux-kernel-build WAF from ~2.2 to ~1.0 in about 10 minutes of config.
- Solidigm's multiple reclaim group FDP turns ~8 reclaim unit handles into up to 1,024 placement IDs (one RG per die), cutting a mixed hyperscale workload's WAF from 3.06 (block mode) to 2.6 (single RG) to 1.01 (simulation) and 1.07 on a prototype SSD, while also tuning bandwidth and QoS.
- ScaleFlux proposes managing FDP overprovisioning per reclaim unit (fairly and prioritized grants), mixing SLC/TLC media in one device, and isolating CPU, SRAM write buffer, and flash command queue resources for more predictable WAF and performance.
Sign in to read the full article.
Sign In