David Heinemeier HanssonandClaude Fable 5.1
dae1f4bf0e
The kernel leaves NVMe on none and everything else on mq-deadline. Neither bounds latency once the device queue fills, so a large build, copy, or package upgrade can make the desktop sluggish while reads wait behind a wall of writes. Kyber keeps separate read and sync-write queues and throttles the depth it submits to hit a 2ms read target, which keeps interactive reads flowing under heavy writes at negligible CPU cost. The trade is a small ceiling on peak throughput on very fast devices, which matters for a storage server chasing IOPS but not for a desktop. Ships as a package-owned udev rule in /etc so it applies at boot and on hot-plug. Existing installs pick it up at the next boot. zram is left alone since it has nothing to schedule. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
7 lines
473 B
Plaintext
7 lines
473 B
Plaintext
# Kyber targets read latency (2ms) and throttles writes to hold it, so the
|
|
# desktop stays responsive under a large build, copy, or package upgrade.
|
|
# The kernel default is none for NVMe and mq-deadline for everything else,
|
|
# neither of which bounds latency once the device queue fills. Applies to real
|
|
# disks only; zram is memory and has nothing to schedule.
|
|
ACTION=="add|change", KERNEL=="nvme[0-9]*n[0-9]*|sd[a-z]*|mmcblk[0-9]*|vd[a-z]*", ATTR{queue/scheduler}="kyber"
|