ionice — Set I/O Scheduling Priority

ionice sets or shows a process's I/O scheduling class and priority — the disk equivalent of nice. Typical use: keep backups, rsync, find, or log compression from making the rest of the system sluggish. It pairs naturally with nice/renice for CPU, and with rsync backup jobs run from cron, so they don't compete with foreground work.

Syntax¶

ionice [-c class] [-n level] [-t] -p PID...      # change running processes
ionice [-c class] [-n level] [-t] -P PGID...     # process groups
ionice [-c class] [-n level] [-t] -u UID...      # all of a user's processes
ionice [-c class] [-n level] [-t] COMMAND [ARGS] # start a command
ionice -p PID                                    # show current setting

Classes¶

`-c`	Class	Behaviour
`0`	none	Default. Priority is derived from the CPU nice value: `(nice + 20) / 5` within best-effort
`1`	realtime	Always served first. Levels 0–7. Root only — can starve everything else, including the system
`2`	best-effort	Normal class. Levels `0` (highest) – `7` (lowest), default `4`
`3`	idle	Only gets disk time when nobody else wants it. No level. No root needed

-n sets the level (0–7) for classes 1 and 2. -t ignores failures to set the class (useful in scripts).

Examples¶

# Run a backup only when the disk is otherwise idle
ionice -c3 tar czf /backup/home.tar.gz /home

# Lowest best-effort priority, combined with low CPU priority
nice -n 19 ionice -c2 -n7 rsync -a /data/ /mnt/backup/

# Check a process
ionice -p 1234
# best-effort: prio 4

# Demote a running process
sudo ionice -c3 -p 1234

# Demote everything a user is running
sudo ionice -c2 -n7 -u builduser

# Give a critical database real-time I/O (careful)
sudo ionice -c1 -n0 -p $(pgrep -o postgres)

When it actually works¶

ionice only sets a hint; the block I/O scheduler decides whether to honour it.

cat /sys/block/sda/queue/scheduler
# none mq-deadline kyber [bfq]

Scheduler	Honours ionice?
`bfq`	Yes — classes and levels
`mq-deadline`	Partially — separates realtime / best-effort / idle classes; levels within a class are mostly ignored
`kyber`, `none`	No

NVMe devices default to none, so ionice usually does nothing on them. Switch to bfq if you need it (at some throughput cost):

echo bfq | sudo tee /sys/block/nvme0n1/queue/scheduler

The old cfq scheduler that ionice was designed around was removed in kernel 5.0.

The buffered-write caveat¶

Most writes go into the page cache and are flushed later by kernel writeback threads — not by your process. The I/O priority mostly applies to reads and direct/synchronous writes. An ionice -c3 process doing heavy buffered writes can still saturate the disk.

For real write throttling, use cgroup v2 limits instead (below).

Better option on systemd systems: cgroups¶

The cgroup v2 io controller works regardless of scheduler for limits, and accounts for writeback:

# One-off command with lower I/O weight (default 100, range 1–10000)
sudo systemd-run --scope -p IOWeight=10 rsync -a /data/ /mnt/backup/

# Hard bandwidth cap
sudo systemd-run --scope -p "IOWriteBandwidthMax=/dev/sda 20M" dd if=/dev/zero of=/tmp/big bs=1M count=2000

In a unit file:

[Service]
IOSchedulingClass=idle          # ionice equivalent (best-effort, realtime, idle)
IOSchedulingPriority=7          # 0-7
IOWeight=10                     # cgroup weight (BFQ honours it)
IOReadBandwidthMax=/dev/sda 50M

IOWeight relies on BFQ for proportional sharing; the *BandwidthMax / *IOPSMax limits work with any scheduler.

Quick reference¶

Goal	Command
Show priority	`ionice -p PID`
Background job	`ionice -c3 CMD`
Low priority, still progresses	`ionice -c2 -n7 CMD`
Low CPU + I/O	`nice -n19 ionice -c3 CMD`
Check scheduler	`cat /sys/block/DEV/queue/scheduler`
Hard limit instead	`systemd-run --scope -p "IOWriteBandwidthMax=DEV RATE" CMD`