Performance & hardware

Boot time

Boot time is a user-visible metric worth tracking per release:

$ systemd-analyze              # total + kernel/firmware split
$ systemd-analyze blame        # slowest units
$ systemd-analyze critical-chain

Budgets for desktop editions: firmware+kernel under 5s on current hardware, userspace under 15s to a login screen. The live ISO boots slower (USB I/O dominates); compare live boots on the same USB stick across releases, not against installed systems.

The usual offenders in profiles: unnecessary enabled services in airootfs, Plymouth on slow disks, and initramfs images built without compression tuning. Squashfs options (-comp xz -b 1M) trade size for decompression time; they are tuned in profiledef.sh.

Resource budgets

Applications and the live session share a floor: the Parch desktop targets 4 GB RAM on installed systems. Practical rules:

• A first-party app idles under ~150 MB RSS.
• The live session idles under ~1.5 GB with the desktop loaded.
• Anything heavier must justify itself in design review (a store caching package metadata is fine; a resident indexer is not).

$ systemd-cgtop              # who is using what, per service
$ ps -eo rss,comm --sort=-rss | head

Hardware compatibility

VMs hide real-world quirks. Before each stable release, one hardware pass covers:

[ ] One Intel laptop with iGPU
[ ] One AMD desktop with dGPU (or APU)
[ ] One NVIDIA machine (proprietary driver path via Driver Manager)
[ ] Wi-Fi (both a common Intel card and a Broadcom card if available)
[ ] Bluetooth audio pairing
[ ] Suspend and resume
[ ] Function keys: brightness, volume, keyboard backlight

DKMS modules (broadcom-wl, nvidia) are the classic hardware-only failure: they rebuild against the kernel at install time. Test them after a kernel upgrade too, not just on first install.

Filing hardware issues

Attach parchprobe output (the hardware probe tool) plus journalctl -b to hardware bug reports; it captures the PCI/USB inventory and kernel state maintainers need. Reports without logs get slower answers.