Software-defined flash (SDF)
SDF is a flash architecture that exposes individual channels and erase operations to host software while retaining per-channel translation and media management.
Interface: The ASPLOS 2014 design uses 8 MB aligned writes and erases with 8 KB reads. Software chooses channels and reclaims blocks; channel engines retain block mapping, dynamic wear leveling, bad-block management and BCH ECC. [1]
Memory: Removing the device write cache does not mean removing all DRAM: the data path still buffers transfers. Acknowledgment after programming does not establish all-or-nothing recovery for an interrupted write. [1] [2]
Capacity: The reported 99% is device-level raw-flash availability. Protection also relies on replication across racks, whose capacity cost is outside that percentage. Eliminating device garbage collection does not eliminate host compaction. [1]
Workload: High aggregate throughput needs concurrent channel access. With one slice, batch size one and 512 KB random reads, SDF trails the conventional comparison device. Lower write-latency variation also need not mean lower mean latency. [1]
Relations
Citations
- Ouyang et al. SDF: Software-Defined Flash for Web-Scale Internet Storage Systems. ASPLOS 2014, pp. 471-484, sections 2-3.
- Peking University copy of SDF, section 2.2: different wording about write completion.
Further reading
Sources & caveats
- Source synthesis of Baidu's 2014 design and workloads, not a hardware reproduction or current-device benchmark. Status is unknown because present availability was not established.
- Section 2.2 differs between copies: the UTA-hosted PDF says synchronous operations; the PKU-hosted PDF says atomic operations. Neither wording alone supplies an interrupted-write recovery protocol. [1] [2]
- The microbenchmark compares 44 synchronous SDF threads against one asynchronous issuing thread for the other devices, with different I/O paths. This is a system comparison, not a controlled FTL-only ablation. [1]