auto-rules-io

I/O, memory and boot tuning for BigLinux.

Your laptop has one disk, and everybody wants it at the same time: the indexer, the package manager, the backup, the browser you just clicked on. Guess which one you are waiting for. This package makes sure it is the browser that gets through, picks a sane I/O scheduler for each disk, and stops the boot from redoing work pacman already did.

No GUI, no "AI-powered optimizer". Udev rules, systemd drop-ins, sysctl files, two short shell scripts and one small user service, deskbalance, that sleeps until a program starts or waits on the disk.

Contents

• What's in the box
• Install
• I/O scheduler
• Disk latency target for the session
• deskbalance
• Faster boot and login
• Memory and swap
• Btrfs housekeeping
• Odds and ends
• License

What's in the box

File	What it does
etc/udev/rules.d/60-ioschedulers.rules	Picks the I/O scheduler per disk type; sets APM level 127 on rotational disks
usr/lib/systemd/system/user.slice.d/50-io-latency.conf	2 ms disk latency target for every login session
usr/lib/systemd/system/user@.service.d/50-io-latency.conf	Hands the io controller to the user manager
usr/lib/systemd/user/{app,session}.slice.d/50-io-latency.conf	The same target inside the session
usr/bin/deskbalance, usr/lib/systemd/user/deskbalance.service, usr/lib/systemd/user.conf.d/50-deskbalance.conf	Moves the job hogging the disk out of the way when a program starts or waits
usr/lib/systemd/system/ldconfig.service.d/50-pacman.conf	Rebuilds the linker cache at boot only when it is missing
usr/lib/systemd/user/app-snap\x2duserd\x2dautostart@autostart.service.d/50-snapd.conf	Runs snapd's login helper only where snapd is enabled
etc/sysctl.d/91-memory.conf	Swap, cache and writeback settings tuned for zram
etc/sysctl.d/99-games-tweaks.conf	Raises vm.max_map_count for games and turns off core dumps
etc/modprobe.d/disable_snd_itco.conf	Blacklists the Intel chipset watchdog and the PC speaker
usr/bin/big-btrfs-fix-to-timeshift, usr/bin/big-btrfs-balance, usr/lib/systemd/system/big-btrfs-balance.service	Resets the Btrfs default subvolume and runs a light balance after boot

Install

It ships in the BigLinux repositories:

sudo pacman -S auto-rules-io

pacman reloads systemd on install, so the drop-ins are picked up right away; the ldconfig and snapd ones only matter at the next boot and login anyway. The scheduler rules and sysctl files live in /etc, where pacman's hooks do not look, so they apply at the next boot, or now with:

sudo udevadm trigger -c change && sudo sysctl --system

The module blacklist waits for the reboot.

The install script enables big-btrfs-balance.service, and on every upgrade it rewrites one specific set of old Btrfs mount options in /etc/fstab (,space_cache=v2,autodefrag,compress-force=zstd:2,commit=15, with or without ,discard=async,ssd) to ,compress-force=zstd. Other fstab lines are left alone.

I/O scheduler

Disk	Scheduler
Rotational sd* (HDD, USB HDD included)	BFQ
Non-rotational sd* (SATA or USB SSD)	Kyber
mmcblk* (eMMC, SD card)	BFQ
nvme*	Kyber

Anything else (vd*, a 27th sd disk) keeps the kernel's default. Rotational disks also get hdparm -B 127, the highest APM level that still lets the disk spin down, provided hdparm is installed.

Disk latency target for the session

What it does

Every login session gets a 2 ms disk latency target on the disks holding / and /home (IODeviceLatencyTargetSec= on user.slice, and on app.slice and session.slice inside the user manager). While your programs miss it, the kernel's io.latency controller throttles everything around them:

• system.slice and machine.slice: system services, containers and VMs, including work done on your behalf such as package updates and Docker builds;
• the user's background.slice: indexers and other polite-in-theory background jobs.

When the session does no I/O, nothing is throttled and everyone runs at full speed. It works under any scheduler, which is the point: ionice classes only act inside one cgroup under BFQ and not at all under Kyber, and IOWeight shares throughput under BFQ but promises nothing about latency.

2 ms is not magic, just measured: 1 ms is slower on HDDs and 3 ms is slower on NVMe.

Numbers

Cold start in seconds, with a sequential reader hammering the disk in the background:

Disk	Target	LibreOffice	Firefox
NVMe	none	16.5–45	65–71
NVMe	2 ms	1.5–2.4	2.1–3.2
HDD	none	104–111	112–132
HDD	2 ms	37–40	74–80

For reference, alone they start in 0.8 and 1.4 s on the NVMe, and 10.6 and 16.5 s on the HDD. The HDD is still an HDD; this package has limits. The NVMe is a Ryzen 7 5700U laptop under Kyber, the HDD an i3-2375M under BFQ, with the page cache dropped before every start. One machine each, so your mileage will vary.

The fine print

• The bill. On the NVMe, a reader that saturates the disk loses about a quarter of its rate while the session does any I/O. Buffered writes and random reads lose nothing measurable, and nothing measurable is lost on the HDD.
• Writeback. On an HDD the target does not help against buffered writes: their writeback happens outside its control.
• Your background counts too. All I/O from the session counts toward the user.slice target, background.slice included. While an indexer such as Baloo or localsearch misses the target, system services and VMs are throttled even if no program waits. The kernel adds each request's latency to every ancestor group with a target, and background.slice lives inside user.slice.
• Storage it can reach. The target was measured on plain partitions. systemd resolves a LUKS volume, or an LVM volume on a single disk, to the disk under it. On RAID, LVM over several disks, multi-device Btrfs or a network /home, the target cannot reach the disks and does nothing.

Turning it off

user.slice holds every login, SSH and root included. On a server, that means an administrator's pacman -Syu would throttle the databases and VMs, which is rarely what anyone wants at 3 AM. Mask the three targets and reboot (the Delegate=io drop-in is harmless on its own):

sudo mkdir -p /etc/systemd/system/user.slice.d /etc/systemd/user/app.slice.d /etc/systemd/user/session.slice.d
sudo ln -s /dev/null /etc/systemd/system/user.slice.d/50-io-latency.conf
sudo ln -s /dev/null /etc/systemd/user/app.slice.d/50-io-latency.conf
sudo ln -s /dev/null /etc/systemd/user/session.slice.d/50-io-latency.conf

The reboot is not optional: systemd drops the setting but leaves the kernel's target in place. Clearing it by hand (target=0 in io.latency) while it throttles leaves the throttled groups delayed until reboot anyway, because the kernel only clears the delay on the group that had the target.

deskbalance

The latency target throttles background.slice, but a copy started from your file manager or a terminal lands in app.slice, right next to the browser you are waiting for, where the target cannot tell them apart. deskbalance is a user service that fixes that:

• A program starts (a unit appears in app.slice): for the 10 s after the last start (30 s at most, then none for 10 s), background.slice and any app.slice unit burning more than an eighth of the CPUs (at least one) share a quarter of the CPUs (at least one), so the new program keeps whole cores.
• A job hogs the disk (at a start, or whenever app.slice waits on the disk): the busiest job of a program scope that has run for a minute or more (younger ones may still be loading) moves to a scope of its own in background.slice, under the latency target, with room for what it allocates plus 256 MB of page cache, and without swap, so a big copy cannot push your programs out of RAM. "Job" means process group: the cp you started from a terminal moves, the terminal stays.
• The job goes quiet for 10 s: it goes back to app.slice, in a scope named after the one it left.

It blocks in inotify, a PSI trigger and a signalfd, so it does not wake up while nothing starts and nothing waits on the disk; while a job is moved, it looks at it once a second. On SIGTERM it lifts the CPU caps and puts every moved job back, and an instance that replaces a killed one takes over what it left. On a kernel without unprivileged PSI triggers, jobs move only when a program starts.

Measured on three laptops while a job in a program scope, started a minute earlier, copies a 3 GB file in a loop. The page cache is dropped before each start, and each program runs without a window and quits once ready; each laptop ran the programs it has installed. Times in seconds, lowest–highest of 3 starts on the NVMe laptops and 2 on the HDD one (with no load, one start where a single value is shown):

	No load	Copy, deskbalance off	Copy, deskbalance on
Asus Vivobook M1503QA (Ryzen 5 5600H, 40 GB, NVMe)
LibreOffice	0.6	2.1–4.3	0.9–4.2
Firefox	1.0	5.8–7.1	1.5–2.4
GIMP	2.6–2.7	15.3–17.1	3.7–4.3
Blender	1.7	11.2–17.0	2.4–3.7
Acer Aspire A515-45 (Ryzen 7 5700U, 36 GB, NVMe)
LibreOffice	0.6	4.2–5.3	0.7–4.0
Firefox	1.3	6.6–10.3	1.4–1.6
GIMP	2.9	10.1–12.7	3.1–4.8
Blender	2.0	8.3–8.9	3.3–3.9
Xonotic server	7.2	8.9–11.7	7.5–7.6
Lenovo IdeaPad S400 (Core i3-2375M, 4 GB, HDD)
LibreOffice	7.1–7.8	50.3–54.2	8.4–9.4
Firefox	19.3	77.5–87.9	21.5–25.2
GIMP	22.3–22.9	114.8–138.7	27.2–29.6
Xonotic server	15.7	57.7–64.2	19.7–20.4

On the NVMe laptops LibreOffice takes the high end when it is the first program to start once the copy is running: the copy then moves as LibreOffice starts, not before.

At rest it uses 0.6 MB of memory and no wakeups.

The fine print:

• A program that does the heavy I/O itself (a browser download, a game updater) moves as a whole, and comes back in a scope with -deskbalance-… added to its name.
• Any unit that appears in app.slice counts as a program starting, D-Bus-activated and timer-started ones included.
• It does not know which window has the focus, so the program you are using can be the one capped for 10 s if it burns a lot of CPU while another one starts.
• On Plasma, krunner, the polkit agent and the kscreen OSD run in background.slice, so they share that quarter of the CPUs for the 10 s after a program starts.
• It reads each program's disk bytes from its cgroup, so the package turns on I/O accounting for user units (DefaultIOAccounting=yes), which applies as soon as the install reloads each user's systemd instance. deskbalance itself starts when that instance starts again: at the next login with no other session of that user open and lingering off, or after a reboot. It caps busy app.slice units only where something there already enabled the CPU controller; otherwise only background.slice is capped, and its log says so.
• The package enables it for every user when it first installs it. To turn it off for everyone: sudo systemctl --global disable deskbalance (upgrades leave that alone). Like enabling, that applies when each user's instance starts again; a deskbalance already running keeps going until then, or stops at once with systemctl --user stop deskbalance in that user's session. For yourself only: systemctl --user mask --now deskbalance, since a user cannot disable what is enabled globally.

Faster boot and login

Linker cache

ldconfig.service rebuilds /etc/ld.so.cache at boot only when the cache is missing or empty, instead of at the first boot after every update. pacman already runs ldconfig after each transaction, so the boot was redoing it inside sysinit.target, where everything else waits: 8.8–9.9 s cold on an HDD, 0.5 s on an NVMe.

If you add libraries without pacman (a systemd-sysext image, a transaction killed halfway), run sudo ldconfig yourself. You were going to anyway, right?

snapd

snapd's XDG autostart (snap userd --autostart) only starts where snapd is enabled, that is, where /run/snapd.socket exists. With snapd turned off it still ran at every login and achieved nothing in 3.3–4.2 s cold on an HDD.

Memory and swap

91-memory.conf assumes zram (or zswap) for swap:

Setting	Value	Effect
vm.page-cluster	0	Swap in one page at a time; zram compresses pages individually
vm.swappiness	30	Prefer dropping cache over swapping
vm.vfs_cache_pressure	75	Keep directory and inode caches a little longer
vm.watermark_scale_factor	150	Wider gap between watermarks, so kswapd starts earlier
vm.watermark_boost_factor	15000	Kernel default, written down: boost watermarks after fragmentation events
vm.dirty_background_ratio / vm.dirty_ratio	5 / 15	Start writeback sooner
vm.compaction_proactiveness	20	Kernel default, written down: background compaction level
vm.extfrag_threshold	300	Compact rather than reclaim a bit sooner (default 500)
vm.panic_on_oom	0	Kernel default, written down: let the OOM killer pick a victim instead of panicking

vm.min_free_kbytes is not here: the pre-systemd-swap service sets it to 3% of RAM, up to 512 MB.

Btrfs housekeeping

big-btrfs-balance.service runs once per boot, after multi-user.target:

1. big-btrfs-fix-to-timeshift resets the default subvolume to the top-level one (the parent of @), which Timeshift expects.
2. Five minutes later, big-btrfs-balance runs btrfs balance with -dusage=50 -musage=50 on /, and on /home if it is a different filesystem, under nice 19 and ionice -c 3 (the latter only counts on BFQ disks).

Odds and ends

• 99-games-tweaks.conf sets vm.max_map_count to 2147483642, the value SteamOS uses for games, and kernel.core_pattern to |/bin/false, which throws core dumps away, so coredumpctl (and any crash reporter built on it) has nothing to show. The 99- prefix puts it last: sysctl files apply in name order, Manjaro's 10-manjaro.conf sets vm.max_map_count, systemd's 50-coredump.conf sets kernel.core_pattern, and the later name wins. If systemd-coredumpd.service is enabled, its systemd-coredump-register.service sets kernel.core_pattern again after sysctl runs, and core dumps come back.
• disable_snd_itco.conf blacklists iTCO_wdt, the Intel chipset watchdog, which nothing on the desktop arms, and pcspkr, the PC speaker, so nothing beeps at you from the motherboard.

License

GPL, as declared in pkgbuild/PKGBUILD.