Sprinkler
Sprinkler is a research SSD scheduler that combines resource-driven request ordering with flash-request grouping to improve internal resource utilization.
Resource-driven I/O scheduling (RIOS) organizes work by chip, die and plane rather than whole host-request order. Flash-level parallelism Aware Request Over-commitment (FARO) feeds compatible work to a chip early. It prioritizes overlap across dies and planes, breaking ties with requests belonging to the same host I/O. Grouping reduces flash transactions, not application data. [1]
Migration matters: a readdressing callback updates scheduling information when live data move between internal resources. A previously resolved placement is not permanently valid. [1]
Evidence: the evaluation uses cycle-accurate simulation and sixteen traces, comparing FARO alone, RIOS alone and their combination. Better within-chip parallelism alone need not minimize whole-device latency. The 2017 manuscript explicitly adds garbage-collection data to the HPCA 2014 work. [1] [2]
Fairness is a separate objective. FLIN later compares against Sprinkler while accounting for inter-flow interference and host priorities. Aggregate throughput does not establish equal slowdown or a tail-latency bound. [3]
Relations
- related
- SSD controller, NAND plane, FLIN
Limitations
- Simulation results are not a benchmark of current commercial SSDs or evidence that a host can enable this policy on arbitrary firmware. [1]
Citations
- Jung and Kandemir. Sprinkler: Maximizing Resource Utilization in Many-Chip Solid State Disks. arXiv:1705.04627v1, 2017, sections 4-5.
- Penn State publication record: HPCA 2014, pp. 524-535. DOI: 10.1109/HPCA.2014.6835961.
- Tavakkol et al. FLIN. ISCA 2018, sections 6-7. DOI: 10.1109/ISCA.2018.00041.
Sources & caveats
- Primary-source synthesis; no native Sprinkler run, hardware measurement or independent reproduction. Historical paper terminology is not a current NAND-command compatibility guarantee.