D2FQ is a fair-queueing I/O scheduler using the NVMe WRR queue arbitration feature. Different from other conventional fair-queueing block I/O schedulers, D2FQ does not stage I/O requests in the block layer but dispatches them immediately to an SSD. Consequently, D2FQ achieves low I/O latency as well as low CPU consumption.

This repository contains the implementation of D2FQ in the Linux kernel.

Please cite the following paper if you use D2FQ:

D2FQ: Device-Direct Fair Queueing for NVMe SSDs. Jiwon Woo, Minwoo Ahn, Gyusun Lee, and Jinkyu Jeong. Proceedings of the 19th USENIX Conference on File and Storage Technologies (FAST 21).

@inproceedings {264838,
author = {Jiwon Woo and Minwoo Ahn and Gyusun Lee and Jinkyu Jeong},
title = {D2FQ: Device-Direct Fair Queueing for NVMe SSDs},
booktitle = {19th {USENIX} Conference on File and Storage Technologies ({FAST} 21)},
year = {2021},
isbn = {978-1-939133-20-5},
pages = {403--415},
url = {https://www.usenix.org/conference/fast21/presentation/woo},
publisher = {{USENIX} Association},
month = feb,
}

1. Compile the kernel with the config CONFIG_IOSCHED_D2FQ enabled. If your NVMe device has a plenty of NVMe SQs, you can also enable CONFIG_IOSCHED_D2FQ_MULTISQ to allow each core to have isolated three SQs (low, med, high). Without this config, three adjacent cores share three SQs (low, med, high). For more information, please refer to the paper.

2. To use D2FQ, a file should be opened with a flag O_D2FQ which is defined in uapi/asm-generic/fcntl.h

3. To enable D2FQ, execute this command echo 1 > /sys/block/path/to/nvme/queue/d2fq_en