Mastering Linux I/O Schedulers: When to Use CFQ, Deadline, or Noop

Linux I/O Scheduler Overview

The I/O scheduler resides in the Linux kernel’s I/O scheduling layer, which is one of seven layers in the overall I/O stack:

VFS layer – virtual file system abstraction.

File system layer – specific file‑system implementations.

Page cache layer – caching of pages.

Block layer – generic block‑device abstraction.

I/O scheduler layer – decides the order of block‑device requests.

Block device driver layer – hardware‑specific driver interface.

Block device layer – the physical device itself.

The scheduler’s goal is to improve overall block‑device performance, especially for mechanical disks where seek time dominates.

1. CFQ – Completely Fair Queueing

CFQ is the default scheduler for most general‑purpose workloads. It creates a separate queue for each process and allocates I/O time slices to ensure fairness.

CFQ also supports priority classes (RT, BE, IDLE) and per‑process I/O priority settings.

Key data structures:

cfq_data – global scheduler state.

cfq_group – represents a cgroup, stored in a red‑black tree keyed by

vdisktime

(total I/O time used).

service_tree – seven trees per group handling RT, BE, and IDLE requests for read/write and async operations.

CFQ parameters (found under

/sys/class/block/<em>dev</em>/queue/iosched/

) include:

back_seek_max

and

back_seek_penalty

– control backward seeks.

fifo_expire_async

/

fifo_expire_sync

– timeout for async/sync requests.

slice_idle

and

group_idle

– idle wait times to improve sequential I/O.

low_latency

and

target_latency

– enable low‑latency mode.

quantum

,

slice_sync

,

slice_async

,

slice_async_rq

– control request batch sizes and time slices.

CFQ also integrates with cgroup blkio control to enforce I/O quotas per group.

CFQ Design Highlights

Each process’s queue is placed in a red‑black tree ordered by its accumulated service time; the scheduler always selects the queue with the smallest

vdisktime

, ensuring fairness across processes and cgroups.

2. Deadline Scheduler

Deadline is simpler and focuses on maximizing throughput while preventing request starvation.

It maintains four queues: read‑sort, write‑sort (ordered by sector), and read‑fifo, write‑fifo (ordered by request age).

When dequeuing, the scheduler prefers reads, checks for any request that has exceeded its deadline (

read_expire

or

write_expire

), and processes it to avoid starvation; otherwise it serves the next request in sector order.

Adjustable parameters include:

read_expire

/

write_expire

– request timeout in milliseconds.

fifo_batch

– number of requests processed per batch.

writes_starved

– threshold for when writes become eligible.

front_merges

– enable/disable forward merges.

3. Noop Scheduler

Noop implements a simple FIFO queue with minimal merging; it is ideal for SSDs where complex scheduling provides no benefit.

Choosing the Right Scheduler

• CFQ : General‑purpose, desktop, and server workloads where fairness among processes is important.

• Deadline : Workloads with heavy I/O pressure on a few processes (e.g., databases) where throughput and bounded latency are critical.

• Noop : SSD or other devices with negligible seek time; the simplest scheduler yields the best performance.

Images illustrating the I/O stack and CFQ data structures: