ScalaAFA: Constructing User-Space All-Flash Array Engine with Holistic Designs
summary
ScalaAFA is a user-space all‑flash array engine implemented on SPDK and evaluated with FAMU-emulated PCIe5 SSDs that offload replication‑to‑striping conversion to drive controllers via custom NVMe commands and the SSD XOR engine. It replaces locks with message-passing permission management, stores frequently updated maps in SSD out‑of‑band metadata, and uses user-space threads plus offload conversion to SSDs to cut CPU overhead. Results show up to 2.5× write throughput and about 52.7% average write latency reduction, reaching near‑ideal throughput with a single host thread.
AI-written summary · about
chapters (14)
- 0:00ScalaAFA project intro
- 1:28Two-phase AFA design challenges
- 3:15High-speed SSD bottlenecks
- 5:23Two-phase AFA beyond parity
- 6:05ScalaAFA holistic design
- 7:27Storage space abstraction
- 8:50Concurrency with space manager
- 9:53Offloading to SSD controllers
- 11:23Mapping table updates
- 13:09Evaluation setup and comparison
- 14:34Microbenchmark results
- 16:21Real workload improvements
- 17:05RocksDB end-to-end eval
- 17:27Test results and conclusion
