Evolution and Mechanisms of Linux I/O Schedulers (Kyber, MQ‑Deadline, BFQ)

The article introduces the concept of I/O schedulers, explaining that traditional spinning disks required head‑movement optimization, while modern SSDs use logical block addresses (LBA) and therefore rely on QoS‑oriented scheduling.

It lists the historical and current Linux block‑layer schedulers, highlighting the mainline ones (noop, deadline, cfq, mq‑deadline, bfq, kyber) and several experimental or out‑of‑tree schedulers (SIO, VR, Row, Zen, Maple, FIO).

Kernel 6.1 community schedulers

Kyber – introduced in 2017, classifies I/O into types (read, write, discard, other) and controls latency per type. It uses four scheduling domains, each with a bitmap indicating resource usage. The scheduler measures per‑request latency (KYBER_IO_LATENCY) and total latency (KYBER_TOTAL_LATENCY) to adjust queue depth and batch size.

MQ‑Deadline – a multiqueue version of the classic deadline scheduler. Requests are inserted into both a sorted RB‑tree (by sector) and a FIFO queue. Expired requests are dispatched first; otherwise, dispatch follows priority and a 2:1 read‑to‑write ratio, with separate deadlines for reads and writes.

BFQ (Budget Fair Queueing) – derived from CFQ, BFQ allocates a budget (maximum sectors) per thread and aims for bandwidth‑fairness. It tracks a virtual time (vtime) per thread, selects the smallest vtime for dispatch, and dynamically adjusts weights based on thread type (interactive, background, newly created, etc.).

The article also covers cgroup/ionice integration: certain schedulers support I/O priority and cgroup weight settings. Example commands are shown:

usage: ionice [-t] [-c CLASS] [-n LEVEL] [COMMAND...|-p PID]

and a kernel structure example:

struct deadline_data *dd = q->elevator->elevator_data;

It presents performance comparisons between idle and real‑time (RT) bandwidth, and provides configuration examples for blkio weight and io.max.

Finally, the article lists several out‑of‑tree scheduler implementations (VR, SIO, Row, Zen, Maple, etc.) with GitHub links, and cites reference documentation such as the Linux kernel block documentation and BFQ tutorials.