Kyber default (#11653): worse read latency under heavy writes than mq-deadline on a SATA SSD with LUKS

[tyz30007]

4.5 sets kyber on all whole disks (60-omarchy-io-scheduler.rules, from #11653) so that "interactive reads keep flowing while a large build, copy, or package upgrade floods the disk with writes". #11653 checked that the rule applies but didn't include latency numbers, so here is a measurement from a SATA machine, the case #3856 was about.

On this machine kyber didn't protect reads under heavy writes. The median read got slower than with mq-deadline, mean and p99 didn't change, and the writer ran slightly faster.

Setup: Samsung SSD 870 EVO 1TB (SATA), LUKS2 (no_read_workqueue, no_write_workqueue, allow_discards), btrfs (compress=zstd:1,discard=async), AMD A4-7210, linux-omarchy 7.2.8-9, Omarchy 4.5.0rc2.

Test: fio 4K random reads at QD1 (psync, direct=1) from a settled file, while a second job writes 1 MiB blocks sequentially (libaio, iodepth 8, direct=1) to a nodatacow file. Each run is 10 s, with the scheduler switched at runtime. The first four runs went A,B,B,A; the last two mq-deadline runs came later, after I switched the disk back for good.

Scheduler	Read median	Read mean	Read p99	Writer
kyber	528 µs	1194 µs	3621 µs	352 MB/s
mq-deadline	399 µs	1160 µs	3719 µs	336 MB/s
mq-deadline	354 µs	1122 µs	3850 µs	334 MB/s
kyber	815 µs	1287 µs	4080 µs	349 MB/s
mq-deadline (later)	440 µs	1174 µs	3817 µs	347 MB/s
mq-deadline (later)	428 µs	1194 µs	4047 µs	342 MB/s

With no writer running, both gave the same read latency (~193 µs mean, ~1 ms p99).

This is one machine, and I didn't test BFQ (#3856's proposal) or NVMe, where kyber may well do better. If others see the same on SATA, maybe limit the rule to NVMe (KERNEL=="nvme*") and leave SATA disks on the kernel default. For now I'm overriding it locally with a later-numbered rule, as #11653 suggests.

fio commands

fio --name=reader --filename=<settled file> --rw=randread --bs=4k --iodepth=1 --ioengine=psync --direct=1 --runtime=10 --time_based \
    --name=writer --filename=<nodatacow scratch file> --rw=write --bs=1M --iodepth=8 --ioengine=libaio --direct=1 --runtime=10 --time_based
echo mq-deadline | sudo tee /sys/block/sda/queue/scheduler   # or kyber, then repeat

Measured and written up with an AI agent (Claude, via Hermes); every number above comes from real runs on this machine.