D2FQ
D2FQ is a research I/O scheduler that selects NVMe priority queues using per-flow virtual time to reduce host-side scheduling overhead.
Queue selection: completed bytes divided by flow weight advance virtual time. Each submission chooses a high, medium or low queue according to its lead over the global reference, avoiding a separate host staging step. Class weights adapt as contention changes. [2]
Arbitration counts commands, not bytes. Assigning a tenant a fixed queue priority therefore does not itself ensure weighted bandwidth sharing when request sizes or competing flow counts differ. D2FQ uses queue classes dynamically rather than dedicating one class to each tenant. [2]
Tuning: tighter virtual-time thresholds can improve short-term fairness but worsen tail latency. These thresholds are service-accounting distances, not time limits on request completion. [1]
Relations
- related
- FlashFQ, FLIN, NVMe submission queue
Limitations
- Requires usable NVMe weighted round-robin support; the prototype does not use the urgent class. [1]
- Short-term unfairness is evaluated empirically without a theoretical upper bound. [1]
Citations
- Woo et al. D2FQ: Device-Direct Fair Queueing for NVMe SSDs. FAST 2021, pp. 403-415, sections 2.3-4.
- Woo et al. D2FQ, FAST 2021 presentation, slides 7-22.
Further reading
Sources & caveats
- Full-paper and presentation review; no D2FQ execution, hardware measurement or replication.
- Evaluation used Linux 5.3.10 and an 800 GB Z-SSD; the NVMe-oF RAM disk tested scalability only, without WRR. [1]